The Equity-Research Report Deck and the Model Walkthrough

Presentation track, branch pr-archetypes (Deck Archetypes) · Node PR5.03 · ~10 focused hours · Mastery gate ≥ 85%

An analyst finishes forty minutes on a mid-cap speciality chemicals company. The target is ₹1,304 against a traded price of ₹1,009, which is 29.2% of upside, and behind it sit a five-year model and eleven exhibits. The portfolio manager asks one question. What does the market believe that you think is wrong?

There is no page that answers it. There is a page on capacity, a page on margin, a page showing the discounted-cash-flow output and a table of comparable multiples. Every number is arithmetically correct. Not one of them says what the traded price already pays for, so nothing in the deck says where the 29.2% is supposed to come from.

The gap that opened in that room is the subject of the next ten hours. A research deck is judged on something a spreadsheet cannot supply on its own: whether a reader can see the difference between your view and the view already embedded in the price, and can interrogate that difference before the meeting ends.

Two things follow, and they are the reason this archetype is harder than the board deck or the consulting readout. The first is that your evidence is a model, and the model is auditable. Somebody in the room will ask to see a cell. If the assumption sheet is missing, or a tax rate is typed inside a formula where nobody can see it, the conversation stops being about the company and becomes about whether you are careful. The second is that the deck has to survive being flipped out of order. Research decks are read alone, on a train, three weeks after the meeting, by a person who was not there and who starts on page seven.

The consequence for craft is specific. Every page carries its claim in the title, its evidence in one exhibit and its provenance in a source line, so that a page removed from its neighbours still says something true. That is a higher bar than a presented deck needs, and it is why research decks are denser than the deck-design literature usually recommends.

There is one further constraint that no other archetype in this branch carries. A research deck is a regulated document in most jurisdictions where it is distributed, and the furniture that makes it compliant is not optional decoration to be squeezed onto the last page at midnight. Ratings definitions, the analyst certification, the conflicts disclosure and the statement of valuation method sit in known places because a reader is entitled to find them without asking.

What follows teaches five artefacts and one performance. The artefacts are the thesis page, the valuation bridge, the scenario page, the risks page and the assumption sheet. The performance is the live model walkthrough, which is nine minutes long when it is done in the right order and twenty-two when it is not.

Learning objectives

You can:

1. Write a thesis page that states the claim, quantifies the variant perception as a difference between two sets of drivers, and names the observable that would prove you wrong.
2. Reverse-engineer the market case from a traded price, express it as a driver set, and use it as the starting point of every valuation exhibit in the deck.
3. Build a valuation bridge from price to target one driver at a time, choose between the sequential and the independent construction, and quantify the interaction term that the independent construction leaves unexplained.
4. Build a scenario page carrying three coherent driver sets, their probabilities, their triggers and a probability-weighted value, and compute the reward-to-risk ratio the page implies.
5. Write a risks page in which every risk names its observable, its effect on value in currency per share and the date by which you would know.
6. Choose the exhibit set a research argument needs, identify the three exhibits that are weak by default, and apply the specific repair each one takes.
7. Build a one-page assumption sheet from a three-statement model, classify every input as reported, policy or estimate, and run the three integrity checks before any slide is built.
8. Walk a model live in the output-first order, in nine minutes, with the four predictable questions answered before they are asked.
9. Show sensitivity honestly, choosing among a two-way grid, a tornado and a scenario set, and stating what a trimmed grid removed and why.
10. Source and date every number on every exhibit, classify it as reported, computed, estimated or market, and compute the coverage share the deck currently achieves.

Prerequisites & connections

Builds on. The board deck and the investor pitch in PR5.02, which own the decisions-first structure, the pre-read and the density that a reading artefact can carry. The archetype idea itself, the action-title body page and the appendix contract come from PR5.01. From the chart branch you need PR2.01 for chart-form choice and PR2.03 for the action title and small multiples, since every exhibit here gets scored against that grammar. On the finance side, you build the model whose output this deck presents in M3.08 and M3.09, and M3.05 teaches the reverse-DCF technique that recovers the market case from a price.

Feeds forward. PR6.01 takes the walkthrough and teaches the Q&A around it, including the question you can't answer. PR7.01 is the capstone, where you choose an archetype under time pressure from an unseen brief, and the research deck is one of the four candidates.

Four neighbours own material that touches this one closely, and the boundaries are drawn on purpose so that none of it gets rebuilt here. M10.03 owns the written analyst deliverables: the one-page memo, the full research report and the exhibit-design drill as applied to written work. Here you get the deck and the live walkthrough. They're different artefacts with a different failure mode, since nobody reading a report can interrupt it. M3.08 and M3.09 own model construction, and nothing below teaches you to build a three-statement model or a DCF. The model in the worked examples comes complete, and the skill you practise is presenting it and defending it. ES1.03 owns the conduct and disclosure obligations themselves, including the advice perimeter and what separates research from a recommendation to a specific client. The compliance section here covers only where the furniture goes on the page and what it doesn't promise. PR5.02 owns the investor pitch, so the market-sizing craft and the ask belong there.

Core teaching content

1. What a research deck must carry that a memo does not

A memo is linear. The reader starts at the top, moves down at a pace they control, and the argument is carried by sentences that can qualify each other across paragraphs. If a claim on page two needs the caveat on page three, prose can hold the two together.

A deck gives up all three of those properties. Your reader is either in a room with you, where you can be interrupted, or opening the file alone six weeks later at wherever the scroll bar dropped them. Sentences are short and page breaks cut them apart, so a caveat two pages away is a caveat nobody read. Prose continuity gets replaced by redundancy of a particular kind: each page states its own claim, shows its own evidence and names its own source.

Three jobs set a research deck apart from the other deck archetypes.

It has to survive being flipped out of order. A portfolio manager who wants the scenario page will go straight to it, and that page must make sense without the eight pages before it. In practice the target price and the traded price appear on every valuation page, front to back. And every chart's title carries its claim, so a reader who lands there cold still gets the point.

It has to support a live challenge to the model. Nobody challenges a memo in real time. In a research meeting somebody opens the model, clicks a cell and reads the formula bar aloud. So the deck needs a page whose only job is to make the model legible before anyone opens it. That page is the assumption sheet.

It has to hold the regulatory furniture where a reader can find it. A memo circulated internally may not carry any. A research deck distributed outside the firm carries the analyst certification, the ratings definitions and distribution, the conflicts disclosure and a valuation-and-risks statement. Where each one goes is a convention, so nobody has to hunt.

What this does to page count surprises people who come from presented decks, so here it is in numbers. The first draft of the Vindhya deck in the worked examples runs 34 pages. After the cut the front section is 11 pages, which is 32.4% of the draft. Nineteen pages move to the appendix and four are deleted outright, so the appendix runs at 1.73 pages for every page at the front. A consulting readout with a 19-page appendix would look bloated. In research it's normal, because the appendix is where the challenge gets answered.

4. Scenarios on one page, with probabilities and triggers

A scenario page carries three coherent worlds. Three arithmetic sensitivities won't do. Coherent means the drivers move together in a way a person could describe. In the bear case the mix shift stalls and pricing pressure arrives, so growth, margin and capital intensity all deteriorate at once and the discount rate rises with the risk.

Here are the three Vindhya cases and what they're made of:

Four things then have to appear on the page, and each of them is a number.

The probability-weighted value is ₹1,256.33. That's ₹47.58 below the base case, because the bear case is 25% likely and ₹610.84 below base, while the bull is 20% likely and ₹525.64 above. The weighted value is still 24.5% above the traded price, so the recommendation survives. It just survives with less room than the base-case headline suggests.

The trigger for each case is the observable that tells the reader which world they're in. A scenario without a trigger is a wish. For the bear case the trigger is two consecutive quarters of gross margin below 30% with the input-cost index rising. For the bull it's a pharma-intermediates share above 35% together with a capacity announcement in performance chemicals.

The reward-to-risk figure is the uncomfortable one, and it belongs on the page. Upside from ₹1,009.22 to the base case is ₹294.70. Downside from the same price to the bear case is ₹316.14. The ratio is 0.93, so the analyst is asking a reader to accept slightly more downside than upside in exchange for a 55% probability that the base case is the right world. You can defend that position. It's just a different position from the one the headline 29.2% implies on its own. Leave the ratio off the page and it doesn't go away. It becomes the first question in the room, and the one you didn't prepare for.

The weighting sensitivity guards against the objection that you picked the probabilities to produce the answer. Move the bear case from 0.25 to 0.40 and take the difference out of the base case, and the weighted value falls to ₹1,164.71, which is 7.3% lower. The recommendation still holds. Push the bear probability to 57.9% and the weighted value equals the traded price. That number tells a reader how far wrong you can be before the call fails.

For the chart, three bars with the traded price drawn as a reference line are all you need. The order bear, base, bull is intrinsic, so don't sort it by value. Probabilities go on the bars as labels. A pie of the three probabilities adds nothing, and it breaks the rule on parts anyway.

5. The risks page that is not a disclaimer

Most risks pages are lists of nouns. Competition. Regulation. Execution. Currency. Input costs. Each one's true and none of them is useful. Together they work as insurance for the analyst, and they tell the reader nothing.

A risk becomes information when it carries four things: the mechanism, the observable that would confirm it's happening, the effect on value in currency per share, and the date by which you'd know. Written that way, a risks page is the most re-read page in the deck, because it's the only page that tells a reader what to watch.

Every number in that table is already in the deck, taken from the bridge bars and the scenario page. That's deliberate. A risks page that introduces new quantities is a page nobody can check.

One more rule belongs here, and it's the one that separates research from advocacy. If the deck contains a finding that argues against the recommendation, it goes on this page, out of the appendix where it would be easier to hide. The Cascade Instruments example later carries such a finding. Its discounted-cash-flow value is $28.43 per share against a peer-multiple value of $45.59, a 37.6% disagreement between two methods in the same deck. Bury that and you haven't removed it, since any reader who divides enterprise value by EBITDA will find it in ninety seconds. Put it on the risks page with your reason for preferring one method, and the rest of the deck becomes credible.

6. The exhibit set, and the three exhibits that are always weak

A research deck's front section carries six exhibits. Not four, because the argument needs a driver chart and a decomposition chart to stand up. Not twelve, because at twelve the reader has stopped choosing what to look at and has started skimming.

The six that carry a research argument:

1. The variant-perception table. Market case against analyst case, driver by driver, in the same units. It's a table because the reader wants to compare ten specific numbers, and that's what tables are for.
2. The valuation bridge. Price to target, one bar per driver, with the interaction handled in the open.
3. The scenario bar with the price line. Three bars, probabilities as labels, traded price drawn across.
4. The driver chart that proves the variant perception. One chart whose only job is to show that the thing you claim is happening is visible in reported data. For Vindhya this is gross margin against the input-cost index over eight quarters.
5. The decomposition exhibit. Where growth actually came from, since a headline growth rate hides which part of the business produced it. For Vindhya, pharma intermediates and performance chemicals together supplied 92.5% of the ₹840 crore of revenue added over two years, while agrochemical intermediates supplied 2.5%.
6. The model output summary. The valuation result, the assumption sheet reference and the three integrity checks. The walkthrough opens on this page.

Three exhibits appear in almost every research deck and are weak by default. Each has a specific repair, and none of the repairs is deletion.

The first is the unannotated price chart: two years of daily closing prices with a moving average, teaching the reader nothing they couldn't get from any terminal. Annotation repairs it. Mark the four events that moved the stock, label each with the event and the one-day move, and title the chart with what the pattern shows. Annotated, the chart tells the reader what the market has already reacted to, and a variant-perception argument needs to know that.

The second is the peer multiple table with no explanatory variable. Five companies, their EV/EBITDA multiples, a mean and a median, and an implied conclusion that the cheapest one is cheap. Repair it by plotting the multiple against the variable that ought to explain it. Across the five Cascade peers, EV/EBITDA against three-year revenue growth fits a line with an R-squared of 0.9918 and a slope of 1.12 turns per point of growth. Now the table has become a claim. At Cascade's growth rate of 7.0% the peer relationship predicts 16.45 times, and Cascade trades at 13.4 times, a residual of 3.05 turns that growth doesn't explain. A scatter is the right form because the message is a relationship.

The third is the generic market-size slide: a total addressable market in billions, sourced to a consultancy press release, with no path from it to the company's revenue line. To repair it, compute the addressable segment from the company's own disclosure and show the share the company already holds. Then the reader can see what growth rate the market size actually permits.

All six exhibits follow the chart grammar you've used on every chart so far. Bars compare across categories. Lines show change over ordered time. Scatters show relationships. Stacked bars carry part-to-whole across few categories and few periods. A slope chart carries change between exactly two periods. Small multiples replace a dual axis every time. And the standing rules hold: no pie above three parts, no third dimension, no dual axes, one message per exhibit, colour that carries meaning or is neutral, and sorting by value unless the order is intrinsic.

8. Walking a model live, output first

The live walkthrough is a performance with a fixed running order, and the order is the difference between nine minutes and twenty-two.

Your instinct will be the bottom-up tour, and it's the wrong one. It starts at the revenue build, walks the cost lines, arrives at the balance sheet, works through the debt schedule and reaches the valuation output around minute twenty-two. By then the audience has asked four questions out of order and you've scrolled six times, and nobody has seen the number the meeting is about. Against the output-first order it wastes 13 minutes, which is 59.1% of the tour.

The output-first order runs in three blocks of three minutes each.

The first block is the output. Open on the valuation page and state the value per share, the enterprise value, the terminal value share and the three integrity checks. For Cascade that's $28.43 per share, an enterprise value of $1,418.1 million and a terminal value that is 83.3% of enterprise value, against an implied market price of $45.59. Then state the three integrity checks and move on. Once the audience knows the answer, they can steer the rest of the tour to the part they doubt.

The second block covers the three drivers that move the answer. That's three of the 24 inputs, and you don't need the rest yet. Take them from the tornado: the discount rate at 32.3% of the total swing, capital intensity at 21.6% and the EBITDA margin at 20.2%. Show where each one appears on the assumption sheet, say where the estimate came from, and give the value at plus and minus its plausible range.

The third block is mechanics and checks, and only now do you open the model. Show the assumption sheet, show one driver row's formula, show the roll-forward that ties and show the cash sweep. By this point a click into a cell is welcome, because the audience already knows what they're looking for.

Four questions turn up in almost every walkthrough, and having the answers ready is the difference between a tour and an interrogation. Someone will ask for the terminal share of value, for the tax rate and its justification, for the effect of a discount rate 80 basis points higher, and for proof that the cash flow statement ties. For Cascade the four answers are 83.3%, a 25% statutory rate with no permanent differences modelled, $24.92 per share, and yes, to twelve decimal places.

Two rules of conduct during the tour. Don't scroll while you talk, because the audience loses their place and yours. Jump between named ranges instead. And when someone finds an error, say what it is, say what it moves and write it down in front of them. An analyst who says the hardcode moves value by 0.41% and offers to send the corrected file before end of day keeps the meeting. One who explains why it doesn't matter loses it.

10. Sourcing and dating every number

Every number on every exhibit belongs to one of four classes, and the source line says which.

A reported number comes from a filing or an official release, and its source line names the document, the table or page and the date. A computed number is your arithmetic on reported numbers, and its source line names the inputs and the operation. An estimated number is your judgement, and its source line says so and names the basis. A market number is a price, a multiple or a yield, and its source line carries a timestamp, because it was true at a moment and is now approximately true.

Auditing a real deck is sobering. Across the six front exhibits of the Vindhya draft there are 47 distinct numbers, an average of 7.8 per exhibit. Thirty-one carry a source line, which is 66.0%. Sixteen don't, and that's 34.0% of everything the reader is asked to accept. Each of those sixteen is either yours, in which case saying so costs four words, or it's a defect.

The fix is a spreadsheet tab. A heroic memory won't do it. One row per number, with the exhibit, the value, the class, the source and the date. Building it takes about twenty minutes for a front section of six exhibits, and it pays for itself the first time somebody asks in a meeting where a figure came from.

Perishable numbers get the harshest treatment, because they're the ones that will be wrong first. Prices, yields, statutory rates, capacity announcements and regulatory thresholds all move. Each one carries the date it was checked, so a reader six weeks later knows what to re-check and isn't left guessing what to distrust. Vindhya's tax rate of 25.17%, the concessional domestic rate including surcharge and cess, carries such a flag: verify against the current Income Tax Act provisions and the Finance Act in force, checked 8 September 2026.

A public-record rule decides which companies can appear at all. A named, real company appears only with facts that are on the public record, cited to the filing and dated. Where the numbers must tie exactly for teaching, as they do throughout the examples here, the company is synthetic and is labelled synthetic on first mention and in the exhibit source line. Both issuers used below are synthetic, so none of their numbers was observed. Every one of them was generated.

11. The compliance furniture and what it does not do

Five pieces of furniture appear in a distributed research deck, and their placement is conventional enough that a reader knows where to look without being told.

The analyst certification states that the views are the analyst's own and that compensation was not tied to the specific recommendation. In the United States, Regulation Analyst Certification requires it, and FINRA Rule 2241 governs research-analyst conflicts more broadly for member firms. Verify the current text and applicability with the primary rulebooks, checked 8 September 2026. In India, research analysts are registered under the SEBI (Research Analysts) Regulations, 2014, and the registration number and prescribed disclosures appear on the document. Verify against SEBI's current regulations and circulars, checked 8 September 2026.

The ratings definitions and distribution state what the rating means in numbers, such as an expected total return above a stated threshold over twelve months, and what share of the firm's coverage carries each rating. A reader who learns that 78% of a firm's ratings are positive reads the next buy recommendation differently, and that's the reason to publish the distribution.

The conflicts disclosure names the relationships: investment banking, market making, holdings above a threshold, directorships.

The valuation method and risks statement names the method used to derive the target and the principal risks to it. It's a compliance requirement in many jurisdictions, and done properly it's also the same content as the risks page. Write it once and write it well, then point to it.

The not-advice line states that the document is general information and not a personal recommendation, and that suitability has not been assessed for any specific reader.

As for placement, the certification, distribution and conflicts go in a fixed block at the back. The rating definition goes beside the rating on the thesis page, since that's where a reader meets the rating. The not-advice line goes in the footer of every page, because pages get separated.

None of this furniture makes an unsupported claim safe. A valuation-method paragraph won't rescue a target price with no method behind it, and the risks statement won't rescue a risks page of nouns. Furniture discloses. It doesn't substantiate. The distinction matters because a great deal of weak research is written as though the second sentence of a disclaimer could go back and repair the first sentence of the thesis.

12. Working from the two templates

Two shipped templates carry the archetype. The research-deck template TPL-equity-research-deck supplies the eleven-page front section as named slots. Each slot comes with its action-title placeholder already written as a claim, and with the exhibit form it expects. The model-walkthrough template TPL-model-walkthrough supplies the assumption sheet layout, the output block, the integrity-check block and the tour script with its three-minute timings.

A template is a floor, and it can turn into a ceiling. The usual failure is filling every slot because it's there, and you end up with an eleven-page deck for a five-page argument. Two checks prevent it.

The first is the read-only-the-titles test you learned with storylines. Print the eleven titles alone. If they read as a connected argument, each one following from the last, the deck has a storyline. If they read as a table of contents, the slots were filled without anyone building the argument. A slot whose title can't be written as a claim should be empty, and an empty slot means a deleted page.

The second is the appendix test. A page that's evidence for a claim, without being the claim itself, belongs in the appendix, indexed so it can be found in one jump. The Vindhya build produced nineteen appendix pages behind eleven front pages. That's a healthy ratio. It means the front section was cut hard and the work was kept.

13. The seam to the finance galaxy

A research deck presents analysis produced elsewhere, so both rubrics apply at once.

The analysis rubric asks whether the model is right: whether the statements tie, whether the drivers are defensible, whether the terminal value is a value rather than a hope, and whether the comparable set is comparable. The communication rubric asks whether the deck carries the argument: whether the titles make claims, whether each exhibit's form matches its message, whether the sourcing is complete and whether a reader who was not in the room can reconstruct the reasoning.

A clean deck on a wrong model scores zero on the first rubric, and nothing on the second can rescue it. The reverse is also true, and it's more common among people who like spreadsheets. A correct model presented as a wall of numbers fails the second rubric completely, and the analysis inside it changes no decision. Both failures are expensive, and only one of them is visible to the person who committed it.

14. Where the assistants help, and where they will hurt you

An assistant is useful at four jobs in this workflow and dangerous at three, and the boundary is sharp enough to write down.

It's good at title surgery. Give it an exhibit's underlying numbers and its current title, and ask for six titles that state a claim the numbers support. Two will be sharper than yours, and rejecting the other four shows you what you actually think.

It's good at the question list. Paste your eleven titles and ask what a sceptical portfolio manager would ask that the titles don't answer. What comes back is a list of appendix pages you haven't built yet.

It's good at source-line auditing. Paste an exhibit's numbers and ask which ones would need a source line under the four-class scheme, and it'll find the estimates you've been treating as facts.

And it's good at compression, turning a paragraph of reasoning into the one supporting sentence a body page allows, as long as you check every figure afterwards against your own computation.

It's dangerous in three places. It will invent a variant perception, because the shape of one is easy to imitate and the substance requires knowing what the price implies. It will produce numbers with false precision, since a plausible number and a correct number look identical in fluent prose, and a valuation bridge is made entirely of numbers. And it will write compliance furniture that reads correctly and isn't correct for your jurisdiction or your firm. That's the one class of error where the cost falls outside the deck.

{
  "skill": "Auditing a research deck's exhibits for sourcing, chart-form fit and title claims before it leaves your desk",
  "use": "After the deck is drafted and every number is computed by you, paste one exhibit at a time as a text table together with its current title, and ask the assistant to classify each number as reported, computed, estimated or market, to say whether the chart form matches the message the title claims, and to name the questions the exhibit invites but does not answer. Fix what it finds; never let it supply a number or a variant perception.",
  "tools": ["Chat assistants (Claude, ChatGPT, Gemini)", "A spreadsheet holding the source register, one row per number"],
  "prompt": "Here is one exhibit from an equity research deck, as a text table, followed by its title. TABLE: <paste>. TITLE: <paste>. Do five things and nothing else. 1) For each distinct number, classify it as reported, computed, estimated or market, and say what a source line for it would have to name. 2) State the message the title claims, and say whether this chart form can carry that message; if not, name the form that can. 3) Name any number on the exhibit that the title's claim does not need. 4) List the three questions a sceptical reader asks after seeing this exhibit. 5) Do not rewrite my numbers, do not estimate anything, and do not propose a target price.",
  "verify": "Recompute every figure the assistant touches from your own script before anything changes on the slide, and check that the rounding on the slide matches the rounding in the computation behind it. Any number it classifies as reported must be traced to a named document, a page or table reference and a date you can open; if you cannot open it, the number is an estimate and is labelled as one. Where it proposes a different chart form, test the proposal against the corpus grammar yourself rather than accepting the label, since an assistant will happily justify a dual axis. Any compliance wording it produces is discarded and replaced with your firm's approved text.",
  "diy": "You must be able to reverse-engineer a market case from a traded price, build the bridge from that case to your own, and state the interaction term, with no tool beyond a spreadsheet you built. That is what the mastery gate tests and it is what the room tests. An analyst who cannot say what the price implies has no variant perception to defend, and no assistant can supply one, because the input it would need is your judgement about the business.",
  "market": "IN"
}

Common mistakes & how experts think differently

1. Stating a target without stating the market case. The draft says the shares are worth ₹1,304 and stops there. It's tempting because the target is the model's output and feels like the finding. An expert treats the target as half a sentence. A number with nothing to compare it against says nothing about what the analyst knows that others don't. Run the reverse DCF first, which puts the market case in place before the target exists, and write the thesis page from the difference.

1. Independent bridge bars presented as though they summed. Five bars, each answering a clean question, drawn as a waterfall that comes up ₹31.53 short of the target. It's tempting because independent bars are easier to explain one at a time, and the discrepancy is only 10.7%. An expert either uses sequential bars and footnotes the order, or keeps the independent bars and draws the interaction as a labelled sixth bar. Arithmetic that doesn't tie, on a page that's about arithmetic, is the most expensive small error in this archetype.

1. Scenarios that are arithmetic rather than worlds. The bear case is the base case with every driver moved down 10%. No reader can picture that, and no trigger can confirm it. An expert writes the story first, as a single line of plain English, then sets the drivers to match it, then attaches the observable that would tell you the story has started. If you can't recognise a scenario while it's happening, it can't inform a decision.

1. The risks page as insurance. Five nouns, no magnitudes, no dates. It's tempting because it's fast and it feels like covering yourself. An expert notices that this is the most re-read page in the deck and fills it with the four columns a reader can act on. Every magnitude comes from bars that already appear elsewhere in the deck.

1. Hiding the exhibit that disagrees with you. The comparable-multiple view says $45.59 and the discounted-cash-flow view says $28.43, so the comps table slips off to the appendix. It's tempting because the deck reads better without it. An expert puts the disagreement on the risks page with the reason for preferring one method. A reader who finds it later stops believing the pages that agreed with you. And a 37.6% method disagreement is itself a finding about how much of the value rests on the terminal assumption.

1. The 25-cell sensitivity grid. Conditional formatting, five values by five, a range from $21.28 to $41.29. It's tempting because it looks thorough and the spreadsheet produces it in one keystroke. An expert asks what question the exhibit answers and notices that a range of 70.4% of the base case answers none of them. Out it goes, replaced by a tornado that ranks the drivers and a trimmed grid whose exclusions are stated in the footnote.

1. The bottom-up model tour. Start at revenue, walk every line, arrive at the answer in minute twenty-two. It's tempting because that's the order the model was built in, and the order the analyst understands it in. An expert opens on the output, because the audience's questions are all downstream of the answer. A tour that reaches the number last spends its first fifteen minutes on questions the number would have retired.

1. Treating the hardcode as small. A tax rate typed inside one formula moves value per share by 0.41%, so it gets fixed without a word, or not at all. An expert reads the same error as a statement about process, since the reader can't bound how many others exist. Before any model is shown, the expert runs the colour convention, the formula audit and one hand-recomputed year.

1. Compliance furniture as an afterthought. The disclosures are pasted onto the last page the night before, and the rating definition never appears beside the rating. An expert places the definition where the reader meets the rating and keeps the not-advice line in every footer, because pages get separated. The valuation-and-risks statement repeats the risks page's content in full.

1. Filling every template slot. The template has eleven pages, so the deck has eleven pages, three of which restate each other. An expert prints the titles alone and reads them as a paragraph. A slot whose title can't be written as a claim gets deleted, and its evidence moves to the appendix, where somebody who wants it can find it.

Worked examples

Both issuers below are synthetic and are labelled synthetic wherever they're named. Every figure comes from computation, and the Verify block after the last example restates all of them at full precision. That way you can reproduce the exhibits yourself and won't need to copy them.

Worked example 1: The thesis page (India, ₹)

Vindhya Speciality Chemicals (synthetic) reported FY26 revenue of ₹4,200 crore across four segments. It has 4.80 crore shares outstanding and net debt of ₹620 crore. The applicable tax rate is 25.17%, the concessional domestic rate including surcharge and cess (check against the Income Tax Act provisions and the Finance Act in force, checked 8 September 2026). Depreciation runs at 3.2% of revenue and net working capital at 22% of revenue.

The traded price is ₹1,009.22. Your five-year free-cash-flow model returns that price exactly on this driver set: revenue growth of 9%, 8%, 7%, 6% and 5%, EBITDA margin flat at 18.2%, capital expenditure at 5.0% of revenue, a discount rate of 11.9% and terminal growth of 4.5%.

Your own base case uses growth of 11%, 10%, 9%, 8% and 7%, margin rising 18.4%, 18.8% and 19.2% then holding, capital expenditure of 5.0%, 5.0%, 4.8%, 4.5% and 4.5%, a discount rate of 11.7% and terminal growth of 4.8%. It returns ₹1,303.92.

Write the thesis page.

Solution. Step 1, compute the claim. Upside is 1,303.9167 / 1,009.2178 − 1 = 0.292007, so 29.2%, and the gap is ₹294.6988 per share. Enterprise value moves from ₹5,464.25 crore to ₹6,878.80 crore, and the terminal value's share of enterprise value rises from 72.65% to 75.74%.

Step 2, name the differences. The two driver sets differ in five places, and only five. That's a design choice worth defending. Every driver you change is a claim you have to support, so a base case that differs from the market case in eleven places has eleven arguments to win, and it'll lose at least one.

Step 3, write the three blocks. The first draft's version of this page ran four bullets: strong growth outlook, margin expansion potential, attractive valuation versus peers, best-in-class execution. Nothing there is falsifiable, and nothing says what the price already assumes.

The corrected page, as a layout in words:

Step 4, verify by an independent route. The market case must return the traded price to the rupee. That's what makes it a market case. Without it you'd only have an opinion about one. Run the model on the market driver set and you get ₹1,009.2178, identical to the price used in the upside calculation. If those two numbers differ, the reverse DCF hasn't converged and the whole thesis page is built on sand.

Read it. The page claims that a specific, observable mix shift is worth ₹295 a share, and it names the measurement that would end the argument. It never claims the company is good. A reader who disagrees now knows which number to attack.

Five checks close a thesis page. Cover the exhibit and read only the words. A reader must still be able to restate the variant perception in one sentence containing a number. The market case must return the traded price exactly. The driver differences must be countable and few, because each one is an argument to win. The falsifier must be an observable with a threshold and a date. And the rating definition goes beside the rating, since that's where a reader meets it.

Worked example 2: The valuation bridge (India, ₹)

Using the same two driver sets, build the bridge from ₹1,009.22 to ₹1,303.92 and choose its construction.

Solution. Step 1, build both constructions. The sequential bridge applies drivers cumulatively in a stated order. Applying growth, then margin, then capital expenditure, then the discount rate, then terminal growth:

The bars sum to ₹294.6988 and the residual against the gap is zero to ten decimal places.

The independent bridge moves one driver at a time from the market case, so each bar answers what that driver alone is worth: growth ₹39.78, margin ₹95.55, capital expenditure ₹58.16, discount rate ₹32.26, terminal growth ₹37.42. Those sum to ₹263.1664 against a gap of ₹294.6988, leaving ₹31.5324 unexplained, which is 10.6999% of the gap.

Step 2, test the order dependence. Run the sequential bridge in the reverse order, starting with terminal growth. The end value is identical at ₹1,303.9167, with a difference of zero to ten decimals. The individual bars still move, though. Revenue growth is ₹39.7756 applied first and ₹58.1668 applied last, a gap of ₹18.3912, which is 6.24% of the total. Terminal growth moves by ₹10.7784 between the two orders.

Step 3, the wrong version. The first draft drew the five independent bars as a waterfall that appeared to reach the target. It can't. Summed from ₹1,009.22 the independent bars reach ₹1,272.38, so the author had stretched the last bar to close the ₹31.53 gap and said nothing about it.

Step 4, the corrected version. Two exhibits are acceptable. The deck uses the first and puts the second in the appendix.

The front-section exhibit is a sequential waterfall in the order growth, margin, capital expenditure, discount rate and terminal growth. Start and end columns are grey and every bar takes one hue, since all five are positive here. The footnote reads that bars are cumulative in the order shown, and that the appendix carries the one-at-a-time decomposition with its interaction term.

The appendix exhibit is the independent decomposition with a sixth bar labelled interaction at ₹31.53, so it also ends at ₹1,303.92.

Step 5, write the title. Operating drivers supply 39.7756 + 104.1438 + 63.6660 = ₹207.5854, which is 70.4398% of the gap. Model settings supply 38.9128 + 48.2007 = ₹87.1135, which is 29.5602%. The title becomes: seven-tenths of the ₹295 gap is operating performance, not a lower discount rate.

Read it. A bridge doesn't prove the target. It shows which arguments the target depends on, so a bridge dominated by the discount rate should worry its own author. Two bridges of the same gap can look entirely different, and the footnote naming the order is what stops that from being a trick.

Six checks close a valuation bridge. The bars sum to the gap, and where they're independent the interaction is drawn and labelled so that they do. The footnote states the order. The order stays intrinsic and unsorted, because this is the one exhibit in the deck where sorting by value would be an error. Start and end columns read as neutral grey. And the title says which class of driver dominates.

Worked example 3: The scenario page (India, ₹)

Build the scenario page from the three coherent driver sets given earlier: bear at ₹693.08, base at ₹1,303.92, bull at ₹1,829.56, with probabilities of 0.25, 0.55 and 0.20.

Solution. Step 1, the weighted value. 0.25 × 693.0779 + 0.55 × 1,303.9167 + 0.20 × 1,829.5563 = ₹1,256.3349. The three contributions are ₹173.2695, ₹717.1542 and ₹365.9113. The weighted value comes in ₹47.5817 below the base case and 24.4860% above the traded price.

Step 2, the distances. Bear is ₹610.8387 below base and bull is ₹525.6397 above it, so the weighted deviation is 0.25 × (−610.8387) + 0.20 × (+525.6397) = −₹47.5817. That reproduces the step-one figure with a residual of zero to ten decimal places. From the traded price, downside to bear is ₹316.1399 and upside to base is ₹294.6988.

Step 3, reward to risk. 294.6988 / 316.1399 = 0.932179. Bull upside from the price is 81.2846% and bear downside is 31.3252%, and the bull-to-bear spread of ₹1,136.4784 is 1.126 times the traded price.

Step 4, the weighting sensitivity. Move bear to 0.40 and base to 0.40, holding bull at 0.20, and the weighted value falls to ₹1,164.7091, which is 7.2931% lower. Now solve for the bear probability at which the weighted value equals the traded price, holding base and bull in their 55:20 ratio on the remainder. It comes out at 57.9047%.

Step 5, the wrong version. The draft page showed three columns of assumptions with the three values at the bottom, no probabilities, no triggers and no traded price anywhere on the page. A reader couldn't tell whether the analyst thought the bear was likely or merely possible. Nor could they see that the bear case is further from the price than the base case is.

Step 6, the corrected version. Three bars in the order bear, base, bull. That order is intrinsic, so the bars must not be sorted by value. The traded price is drawn as a horizontal reference line at ₹1,009.22. Each bar is labelled with its value and its probability. Below the chart sit three rows, one per scenario, each naming the story in one clause and the trigger in one clause. Beside the chart sits a three-line block: probability-weighted value ₹1,256, reward to risk 0.93, and the bear probability at which the call fails, 57.9%.

The title states the finding: the base case pays ₹295 and the bear case costs ₹316, so this is a 55%-probability call rather than a cheap option.

Read it. A reward-to-risk ratio below 1.0 doesn't mean you abandon the recommendation. It means the position gets sized differently, and that decision belongs to the reader. The page's job is to hand over the number. Hiding it behind an average lets the reader down.

Six checks close a scenario page. Probabilities sum to 1. Each case carries an observable trigger with a threshold. The traded price appears on the chart as a reference line. Bear, base and bull stay in their intrinsic order, unsorted. The probability-weighted value is shown alongside the base case and never replaces it. And the reward-to-risk ratio is stated even when it's unflattering, because a reader who works it out in their head has caught you either way.

Worked example 4: Where the growth came from (India, ₹)

Vindhya's segment revenue, in ₹ crore, for FY24 and FY26: agrochemical intermediates 1,428 and 1,449; pharma intermediates 924 and 1,323; performance chemicals 672 and 1,050; other 336 and 378. Totals are 3,360 and 4,200. Build the exhibit that proves the mix claim in the thesis.

Solution. Step 1, compute the shares and check they tie. FY24 shares are 42.5%, 27.5%, 20.0% and 10.0%, summing to 100.0000000000. FY26 shares are 34.5%, 31.5%, 25.0% and 9.0%, also summing to 100.0000000000. Share changes are −8.0, +4.0, +5.0 and −1.0 percentage points.

Step 2, compute the growth contribution, which is the real message. Total revenue rose ₹840 crore. Agrochemical intermediates contributed ₹21 crore, pharma ₹399 crore, performance chemicals ₹378 crore and other ₹42 crore, and the four add back to ₹840 crore with nothing left over. As shares of the increase, that's 2.5%, 47.5%, 45.0% and 5.0%. Pharma and performance chemicals together supplied 92.5%.

Two-year compound growth rates tell the same story from a different angle: 0.7326% for agrochemicals, 19.6586% for pharma, 25.0000% for performance chemicals and 6.0660% for other, against 11.8034% for the company.

Step 3, the wrong version. The draft used two pies side by side, FY24 and FY26, four slices each. That's three faults, in ascending order of seriousness. First, the chart grammar rules out a pie above three parts, and this one has four. Second, comparing two pies makes the reader hold eight angles in memory. The two segments whose difference matters most, pharma at 113.4 degrees and performance chemicals at 90.0 degrees, differ by 23.4 degrees. That's 6.5 percentage points of share, and it's really hard to read as an angle. Worst of all, neither pie shows the growth contribution, which is the actual claim. A share can rise while a segment shrinks, so share and contribution are different messages.

Step 4, the corrected version. Two exhibits, one at the front and one in the appendix.

The front exhibit is a sorted horizontal bar of contribution to the ₹840 crore of growth, largest first: pharma 47.5%, performance chemicals 45.0%, other 5.0%, agrochemicals 2.5%. Sorting by value is right here, because the order isn't intrinsic. Two bars carry the message colour and two are neutral grey, so the colour carries meaning. The title reads that two segments supplied 92.5% of two years of growth while the largest segment supplied 2.5%.

The appendix exhibit is a slope chart of share, FY24 to FY26, one line per segment, labelled at both ends. A slope chart is the form for change between exactly two periods, and it shows the −8.0 point fall in agrochemicals alongside the +5.0 and +4.0 point gains without asking anyone to compare angles.

Step 5, verify by an independent route. Contribution shares must sum to 100%: 2.5 + 47.5 + 45.0 + 5.0 = 100.0. Segment increases must sum to the total increase, and the residual is 0.0000000000. Run both, because a segment table with a restated prior year will fail them, and the failure is the finding.

Read it. The company's headline two-year growth of 11.8% a year is an average over two businesses growing at 19.7% and 25.0% and one growing at 0.7%. That average is the number in the press release, and it's the least informative number available.

Five checks close a decomposition exhibit. Where the claim is about change, the exhibit shows contribution to change. The contributions sum to the total, and any residual is a finding, however much it looks like rounding. The bar is sorted by value, since the order isn't intrinsic. Colour marks the two segments the title names and nothing else. And the title carries the 92.5%. The word mix on its own says nothing.

Worked example 5: Gross margin against the input-cost index (India, ₹)

Eight quarters of Vindhya's gross margin, in percent, from Q1FY25: 31.2, 30.6, 29.8, 29.1, 29.4, 30.1, 31.0, 32.2. The input-cost index over the same quarters, rebased to 100 at the start: 100.0, 104.5, 110.2, 114.8, 112.3, 107.6, 102.9, 98.4. Build the driver chart that proves the margin claim.

Solution. Step 1, compute what the chart has to say. The margin trough is 29.10% in Q4FY25 and the latest reading is 32.20%, a recovery of 3.10 percentage points. Over the full eight quarters margin is up 1.00 point. The index peaked at 114.80 in the same quarter as the margin trough and has fallen 14.2857% from that peak, ending 1.60% below where it started.

The same-quarter correlation between margin and index is −0.978364. Lagged one quarter it's −0.596822, and lagged two quarters it turns positive at +0.230242. That's the evidence that the pass-through is fast rather than delayed.

Step 2, the wrong version, and why it is wrong in a way that is hard to see. The draft plotted margin on the left axis and the index on the right, two lines, one chart. The exhibit looked convincing, and the point where the two lines crossed seemed to mark the turn.

That crossing point is an artefact of a choice nobody can see. With the left axis running 28 to 33 and the right axis 95 to 120, the lines cross at quarter index 1.3608. Change the right axis to 90 to 130 and they cross at 1.5207. Change it to 80 to 160 and they cross at 1.9243. The crossing moves by 0.5635 of a quarter across those three choices, and every one of them is a defensible axis range. The reader sees a fact. The author chose it.

Step 3, the corrected version. Two repairs, and the deck uses both in different places.

For the front section, index both series to 100 at Q1FY25 and plot them on one axis. The margin index ends at 103.2051 and the input index at 98.4000, a spread of 4.8051 index points. Now the divergence is a fact about the data, and the axes have no say in it. With a single axis the reader can measure the gap.

For the appendix, use small multiples: two panels stacked, sharing an x-axis, each with its own y-axis clearly labelled in its own units. The chart grammar prefers small multiples to dual axes every time, and step two just showed you why.

Step 4, the title. Gross margin has recovered 3.1 points from the Q4FY25 trough as the input index fell 14.3% from its peak, with a same-quarter correlation of −0.98.

Read it. A correlation of −0.98 across eight observations is a strong relationship and a weak proof, since eight points and one input cycle can't separate pass-through from mix. Stated fairly, the claim is that margin recovery so far is consistent with input costs alone. The mix argument needs the segment exhibit to stand on. A driver chart that overclaims invites the one question the analyst can't answer.

Four checks close a driver chart. Look for a second y-axis first, and where there is one, delete it and choose between indexing and small multiples. The title states the relationship together with its magnitude. Any correlation is quoted with its sample size. And read the chart once more for a causal claim the data can't support. That's the failure that invites the question you can't answer.

Worked example 6: The assumption sheet and the hidden hardcode (US, $)

Cascade Instruments Inc. (synthetic) has FY0 revenue of $840.0 million and 42.0 million shares. The three-year forecast runs on revenue growth of 8%, 7% and 6%, with cost of goods at 58.0%, 57.5% and 57.0% of revenue and operating expenses at 23.0%, 22.9% and 22.9%. Depreciation and amortisation run at 4.4% of revenue across the forecast, against $37.0 million reported in FY0, and capital expenditure at 5.2%. Working capital is set by receivable days of 52, inventory days of 74 and payable days of 46, with accrued liabilities at 5.7% of revenue. The capital structure is a term loan of $320.0 million at 6.25% amortising $20.0 million a year, a revolver at 7.50% with a $60.0 million minimum cash balance, and cash earning 4.00%, all taxed at 25%. Opening cash is $96.0 million, net property is $410.0 million and goodwill is $265.0 million.

Build the model output page and the assumption sheet, then find what is wrong with the model.

Solution. Step 1, run the integrity checks first. Total assets less liabilities and equity is 0.000000000000 in all four years. The property roll-forward and the equity roll-forward both residual to 0.000000000000 in each forecast year, and the cash flow statement ties to the balance sheet cash line to the same precision in all three. Those three lines go on the output page as text.

Step 2, the output block. Revenue grows from $840.000 million to $1,028.946 million, a compound rate of 6.9969%. EBITDA rises from $159.600 million to $206.818 million and the margin from 19.0% to 20.1%, a gain of 1.10 points. Net debt swings from $224.000 million to negative $31.624 million, a movement of $255.624 million. A company carrying net debt of 1.40 times FY0 EBITDA therefore ends the forecast with net cash.

Unlevered free cash flow is $83.3476 million, $95.0140 million and $105.7748 million. At a 9.4% discount rate and 2.4% terminal growth, the present value of the explicit period is $236.3586 million, the terminal value is $1,547.3336 million and its present value $1,181.7673 million. That present value is 83.3330% of the enterprise value of $1,418.1260 million. Equity value is $1,194.1260 million and value per share is $28.4316.

Step 3, the assumption sheet. Twenty-four inputs on one page, four columns. Seven are reported actuals, five are policy or contractual and twelve are analyst estimates, so 50.0% of the sheet is judgement. The type column is the one a sceptical reader scans first.

Step 4, find the hardcode. The FY2 tax line reads 21% of pre-tax income. The assumption sheet says 25%. The lower rate is typed inside the formula, where no input cell shows it.

Here are its effects, in order of visibility. FY2 pre-tax income is $134.6861 million, so tax should be $33.6715 million. It's $28.2841 million, understated by $5.3874 million. FY2 net income becomes $106.4020 million against a correct $101.0146 million, an overstatement of 5.3333%. The knock-on into FY3 is $0.1616 million of extra net income, from interest on cash that shouldn't exist, and FY3 closing cash is $297.1726 million against a correct $291.6236 million. FY2 unlevered free cash flow is $100.9159 million against $95.0140 million. Three-year cumulative net operating profit after tax is overstated by $5.9019 million. And value per share becomes $28.5490 against $28.4316, higher by $0.1174 or 0.4130%.

Step 5, the wrong version and the corrected version. The draft presented the model as a screenshot of the spreadsheet, four columns of forty rows, shrunk to fit. Nothing on it was legible at reading distance and nothing on it was checkable. The audience did what audiences do with an illegible exhibit and asked for the file.

The corrected page carries three blocks and no screenshot. Its output block gives the six numbers above. Then the assumption block gives the 24 inputs with their type and source, or a pointer to the facing page if the sheet needs a full page. A checks block gives the three integrity lines with their residuals. And the title states the finding: the model produces $28.43 a share, with 83.3% of it in the terminal value.

Read it. A 0.41% valuation effect sounds like nothing. But a reader who finds the hardcode can't bound the number of similar errors, so a defect worth eleven cents a share costs you the entire meeting. This is the reason the colour convention, the formula audit and one hand-recomputed year all happen before the deck is built. Waiting until a client finds something is too late.

Five checks close a model output page. The three integrity checks appear on the page with their residuals. The terminal value share is stated, so nobody has to compute it. No screenshot of a spreadsheet appears anywhere in the deck. Every input on the assumption sheet carries a type and a source. And a stranger can find every chosen number without opening the model, which is the test all four of the others serve.

Worked example 7: Showing sensitivity without drowning the reader (US, $)

Using the Cascade model, produce the sensitivity exhibit for the front section and decide what goes to the appendix.

Solution. Step 1, the two-way grid. Discount rates of 8.4%, 8.9%, 9.4%, 9.9% and 10.4% against terminal growth of 1.4%, 1.9%, 2.4%, 2.9% and 3.4% produce values per share from $21.2792 to $41.2877. The maximum is 1.9403 times the minimum and the range is 70.3741% of the $28.4316 base case.

Step 2, why that grid fails as an exhibit. An exhibit exists to support a claim. A grid whose corners span 70% of the base case supports the claim that the model is sensitive to its own settings. That's true of every discounted cash flow ever built, and it isn't what the page was for.

Step 3, trim it honestly. The middle three by three, discount rates of 8.9% to 9.9% against terminal growth of 1.9% to 2.9%, runs $24.4098 to $33.7902. That range is 32.9930% of the base case and keeps 46.8824% of the full grid's range while removing 16 of the 25 cells. The footnote states which rows and columns were removed and why. A discount rate of 8.4% falls below any defensible cost-of-capital build for this issuer, and terminal growth of 3.4% approaches long-run nominal growth.

Step 4, build the tornado, which is the front-section exhibit. Five drivers, each moved across its plausible range with the others held: discount rate ±80 basis points, capital intensity ±80 basis points of revenue, EBITDA margin ±100 basis points, terminal growth ±50 basis points and revenue growth ±2.0 points a year. The centre is $28.4316 and matches the base discounted cash flow with a difference of zero to ten decimals. That match is your check that the tornado was built on the same model as the headline.

Sorted by swing: discount rate $7.936, capital intensity $5.304, EBITDA margin $4.972, terminal growth $4.316, revenue growth $2.055. The swings total $24.5828, so the shares are 32.28%, 21.57%, 20.23%, 17.56% and 8.36%, and the top two carry 53.8569%.

Step 5, the wrong version. The draft put the 25-cell grid at the front with conditional formatting, green at the top left and red at the bottom right, and no tornado anywhere. It answered a question nobody asked and buried the ranking that would have told the reader where the argument is fragile.

Step 6, the corrected version. The front section gets the tornado, sorted by swing, with the base value marked on the centre line and the range tested named on each bar. Its title says two drivers carry 53.9% of the swing, and neither is revenue growth. The appendix gets the trimmed grid with its exclusion footnote, and the full grid below it for anyone who wants it.

Read it. The tornado also tells you where to work. A 20 basis point cut in the discount rate adds $1.0078 a share against $0.5153 for a full extra point of annual revenue growth, a ratio of 1.96, so the cost-of-capital build gets the next hour. Put the exhibit's limitation on it in one line: drivers are moved one at a time, so correlated moves produce a wider swing than any single bar shows.

Five checks close a sensitivity exhibit. Name the question first. A grid answers what the value is at a combination and a tornado answers which driver matters, and usually only one of those is the page's question. The tornado centre reproduces the headline value exactly, which is how you know it was built on the same model. The tornado is sorted by swing. A trimmed grid carries a footnote naming what was removed and justifying it by the driver's plausible range. And the one-at-a-time limitation is stated on the exhibit.

Worked example 8: The peer exhibit and the finding that argues against you (US, $)

Five peers for Cascade Instruments, with EV/EBITDA, three-year revenue growth and EBITDA margin: Northwind Metrology 14.8 times, 5.2%, 18.4%; Argent Test Systems 17.9 times, 8.1%, 20.6%; Halden Process Controls 12.1 times, 3.4%, 16.9%; Peregrine Sensors 20.4 times, 10.7%, 22.1%; Calder Analytical 15.6 times, 6.3%, 19.2%. All five are synthetic. Cascade trades at 13.4 times FY0 EBITDA, grows at 7.0% and runs a 19.0% margin.

Solution. Step 1, the table version and what it implies. The peer mean is 16.16 times and the median 15.6 times. Cascade at 13.4 times is 2.76 turns below the mean, a 17.0792% discount. A table stops there and leaves the reader to infer that the discount is an opportunity.

Step 2, the scatter version. Plot EV/EBITDA against three-year growth for the five peers. The fitted line has an intercept of 8.5992 turns and a slope of 1.1218 turns per point of growth, with an R-squared of 0.9918 across five observations. Fitted values are 14.432, 17.686, 12.413, 20.602 and 15.666 times, and residuals are +0.368, +0.214, −0.313, −0.202 and −0.066 turns, all small.

At Cascade's 7.0% growth the relationship predicts 16.4517 times. Cascade trades at 13.4, a residual of −3.0517 turns. Growth explains none of the discount to the mean. Cascade grows faster than the peer average of 6.74%, so the fitted line puts it 0.2917 turns above the mean. That makes the unexplained gap wider than the raw 2.76-turn discount suggests.

Step 3, the finding that argues against the recommendation. The discounted cash flow values Cascade at $28.4316 a share. The enterprise value it implies, $1,418.1260 million, is 8.8855 times FY0 EBITDA of $159.600 million. Meanwhile the traded multiple of 13.4 times implies an enterprise value of $2,138.6400 million, equity of $1,914.6400 million and a share price of $45.5867. The discounted cash flow is 37.6318% below the price the multiple implies.

Two methods in the same deck disagree by more than a third. They can't both be right.

Step 4, the wrong version. The draft kept the peer table at the front, showing the 17.1% discount to the mean, and moved the discounted-cash-flow output to the appendix. The deck therefore argued that the stock was cheap on multiples while carrying, unread, the evidence that its own model thought it expensive.

Step 5, the corrected version. The front section carries the scatter with its fitted line, Cascade marked and its residual labelled, titled with the finding that growth explains none of the discount. The risks page carries the method disagreement in one row. The peer relationship implies $45.59 and the discounted cash flow implies $28.43, a 37.6% gap. One clause gives the reason for preferring the cash flow view, and another gives the condition under which the multiple view would win.

Read it. An R-squared of 0.9918 on five points gives you a fitted line. It doesn't give you a law. A residual of three turns may mean the market is wrong, or it may mean the peer set is wrong. Where the terminal value is 83.3% of enterprise value, as it is here, a discounted cash flow is an opinion about the far future wearing a decimal point. The deck's job is to show the reader both opinions and say which one the recommendation rests on.

Four checks close a peer exhibit. The exhibit plots the multiple against the variable that should explain it, so it does more than list multiples. The fitted line's sample size is stated beside its R-squared, because five points can produce almost any R-squared. The subject company is marked and its residual labelled. And where two methods disagree, the disagreement goes on the risks page and stays out of the appendix.

Verify

Every figure printed above, at full precision, in the order it appears.

Vindhya market case: value per share 1,009.2178, enterprise value 5,464.2455, equity 4,844.2455, present value of explicit flows 1,494.2048, present value of terminal 3,970.0408, terminal share 0.726549. Analyst case: 1,303.9167, 6,878.8000, 6,258.8000, 1,669.0857, 5,209.7143, 0.757358. Upside 1,303.9167/1,009.2178 − 1 = 0.292007; gap 294.6988.

Sequential bridge, growth first: +39.7756, +104.1438, +63.6660, +38.9128, +48.2007, running to 1,048.9934, 1,153.1372, 1,216.8032, 1,255.7160, 1,303.9167; sum 294.6988; residual 0.0000000000. Reversed order: +37.4223, +34.6987, +62.4353, +101.9758, +58.1668; end 1,303.9167; difference across orders 0.0000000000; growth-bar gap 18.3912; terminal-growth-bar gap 10.7784. Independent bars: +39.7756, +95.5462, +58.1622, +32.2602, +37.4223; sum 263.1664; interaction +31.5324; interaction share 0.106999. Bar shares of the gap: 13.4970%, 35.3391%, 21.6037%, 13.2043%, 16.3559%; operating bars 70.4398%; model settings 29.5602%; margin plus capex 167.8098 and 56.9428%.

Scenarios. Bear drivers: growth 6/5/4/4/4%, margin 17.0/16.8/16.8/16.8/16.8%, capex 5.2/5.2/5.0/5.0/5.0%, discount 12.5%, terminal 4.0%. Bull drivers: growth 14/13/12/10/9%, margin 19.5/20.0/20.5/20.5/20.5%, capex 5.0/4.8/4.5/4.2/4.2%, discount 11.3%, terminal 5.2%. Values: bear 693.0779, base 1,303.9167, bull 1,829.5563; weighted 1,256.3349; contributions 173.2695, 717.1542, 365.9113; weighted less base −47.5817; weighted upside on price 0.244860. Bear less base −610.8387; bull less base +525.6397; weighted legs −152.7097 and +105.1279, residual −0.0000000000. Downside to bear 316.1399; upside to base 294.6988; upside to bull 820.3385; reward to risk 0.932179; bear downside −31.3252%; bull upside 81.2846%; spread 1,136.4784 and 1.126098 times the price. Reweighted 40/40/20: 1,164.7091, change −91.6258 and −7.2931%. Break-even bear probability 0.579047.

Segments FY24 and FY26 (₹ crore): 1,428.00/1,449.00, 924.00/1,323.00, 672.00/1,050.00, 336.00/378.00; totals 3,360.0000 and 4,200.0000; two-year CAGR 0.118034. Shares FY24 42.5000, 27.5000, 20.0000, 10.0000 summing to 100.0000000000; FY26 34.5000, 31.5000, 25.0000, 9.0000 summing to 100.0000000000; changes −8.0000, +4.0000, +5.0000, −1.0000. Segment CAGRs 0.7326%, 19.6586%, 25.0000%, 6.0660%. Increases 21.0000, 399.0000, 378.0000, 42.0000, summing to 840.0000 with residual 0.0000000000; shares of the increase 2.5000%, 47.5000%, 45.0000%, 5.0000%; two segments 777.0000 and 92.5000%. FY26 pie angles 124.2000, 113.4000, 90.0000, 32.4000 degrees; the pharma-to-performance angle difference 23.4000 degrees on a share difference of 6.5000 points.

Quarterly margins 31.2, 30.6, 29.8, 29.1, 29.4, 30.1, 31.0, 32.2; index 100.0, 104.5, 110.2, 114.8, 112.3, 107.6, 102.9, 98.4. Trough 29.10 in Q4FY25; latest 32.20; recovery +3.1000; eight-quarter change +1.0000. Index peak 114.80; fall from peak −14.2857%; eight-quarter change −1.6000%. Correlations −0.978364, −0.596822, +0.230242. Dual-axis crossings 1.3608, 1.5207, 1.9243; moves 0.1598 and 0.5635. Rebased latest values 103.2051 and 98.4000, spread 4.8051.

Cascade income statement, FY0 to FY3 ($ million): revenue 840.000, 907.200, 970.704, 1,028.946; EBITDA 159.600, 172.368, 190.258, 206.818; EBIT 122.600, 132.451, 147.547, 161.545; pre-tax 106.440, 116.291, 134.686, 152.548; net income 79.830, 87.218, 101.015, 114.411; cash 96.000, 147.228, 212.596, 291.624. EBITDA margin 19.0000%, 19.0000%, 19.6000%, 20.1000%; FY0-to-FY3 revenue CAGR 6.9969%; margin gain +1.1000 points; FY3 net debt −31.6236, swing −255.6236; FY0 net debt to EBITDA 1.403509. All integrity residuals 0.000000000000.

Cascade valuation: unlevered free cash flow 83.3476, 95.0140, 105.7748; present value of explicit 236.3586; terminal value 1,547.3336; its present value 1,181.7673; terminal share 0.833330; enterprise value 1,418.1260; equity 1,194.1260; value per share 28.4316. Hardcode: pre-tax 134.6861; tax 33.6715 against 28.2841, understated 5.3874; net income 101.0146 against 106.4020, overstated 5.3333%; FY3 net income difference 0.1616; FY3 cash 291.6236 against 297.1726; FY2 unlevered flow 95.0140 against 100.9159; cumulative NOPAT difference 5.9019; value per share 28.5490, higher by 0.1174 and 0.4130%; enterprise value higher by 4.9312.

Grid: minimum 21.2792, maximum 41.2877, ratio 1.940281, range 70.3741% of base; trimmed 24.4098 to 33.7902, range 32.9930% of base, keeping 0.468824 of the full range, removing 16 of 25 cells. Tornado centre 28.4316, matching the base case with difference 0.0000000000; swings 7.936, 5.304, 4.972, 4.316, 2.055; total 24.5828; top two 13.2396; shares 0.322824, 0.215746, 0.202261, 0.175587, 0.083582; top two 0.538569.

Peers: fitted values 14.432, 17.686, 12.413, 20.602, 15.666; residuals +0.368, +0.214, −0.313, −0.202, −0.066; mean 16.160000, median 15.6000, mean growth 6.740000; intercept 8.599174, slope 1.121784, R-squared 0.991817. Cascade fitted 16.451664, residual −3.051664; discount to mean −2.760000 turns and −17.0792%; growth effect +0.291664 turns, share −0.105675. FY0 EBITDA 159.6000; discounted-cash-flow multiple 8.885501; enterprise value at 13.4 times 2,138.6400; equity 1,914.6400; price 45.5867; discounted cash flow against it −37.6318%.

Driver-revision comparison on the Cascade model: value at a discount rate 20 basis points lower 29.4394 and 20 higher 27.4798, so the upward move is 1.0078 and the downward 0.9518; value at growth one point higher 28.9468 and one point lower 27.9194, moves of 0.5153 and 0.5121; value at growth two points higher 29.4652, a move of 1.0336; ratios 1.956030 and 0.975081. Bridge residuals: growth-bar order gap as a share of the total 6.2407%; independent bars stacked from the price 1,272.3843 and 89.3001% of the gap; operating bars 207.5854 and model settings 87.1135; reweighted 40/40/20 value against the price 15.4071%; analyst-case margin gain over the market case by year three 1.0000 point.

Deck arithmetic: draft 34 pages, front 11, appendix 19, deleted 4, checking to 34; front share 32.3529%; appendix per front page 1.727273. Source register: 47 numbers, 31 sourced, 16 unsourced, 65.9574% and 34.0426%, 7.833333 numbers per exhibit. Assumption sheet 24 inputs, 7 actual, 5 policy, 12 estimate, residual 0, estimates 50.0000%. Tour 22.0 minutes bottom up against 9.0 output first, saving 13.0 minutes and 59.0909%, in three blocks of 3.0 minutes with residual 0.0.

Practice set

Work each problem before you read its solution. You'll need a calculator. Numeric answers within 2% of the stated figure count as correct unless a problem says otherwise. Where a problem asks for a title or a page, the test is whether a reader could disagree with what you wrote.

P1 (guided). The claim line (India). Vindhya trades at ₹1,009.2178 and your base case returns ₹1,303.9167. State the upside and the rupee gap, then write the rating strip as it would appear on the thesis page.

Solution. Upside is 1,303.9167 / 1,009.2178 − 1 = 0.292007, so 29.2%, and the gap is 1,303.9167 − 1,009.2178 = ₹294.6988 per share. The strip reads: BUY, target ₹1,304, price ₹1,009, upside 29.2%, twelve-month horizon, with the rating definition beside it stating what BUY means as an expected total return threshold. The common wrong answer leaves out the price. Then a reader can't check the upside without leaving the page, and the page has to survive being read alone.

P2 (guided). The terminal share (US). Cascade's enterprise value is $1,418.1260 million, of which the present value of the terminal value is $1,181.7673 million. Compute the terminal share and say what you do with it on the output page.

Solution. 1,181.7673 / 1,418.1260 = 0.833330, so 83.3%. It goes on the output page as a stated line where nobody can miss it, because it's the first question a sceptical reader asks and it tells them where to push. A model with 83.3% of its value beyond the explicit forecast is an opinion about the terminal assumptions. So the discount rate and terminal growth deserve the argument, and the revenue build doesn't. The tornado agrees: those two drivers carry 32.28% and 17.56% of the total swing.

P3. The missing bar (India). The five independent bridge bars are ₹39.7756, ₹95.5462, ₹58.1622, ₹32.2602 and ₹37.4223, and the gap from price to target is ₹294.6988. Compute what is missing, express it as a share of the gap, and say what you draw.

Solution. The bars sum to 263.1664, so 294.6988 − 263.1664 = ₹31.5324 is unexplained, which is 31.5324 / 294.6988 = 0.106999, or 10.7% of the gap. That residual is the interaction between drivers, since a higher margin applied to a larger revenue base is worth more than the two effects added separately. You draw it as a sixth bar labelled interaction, so the waterfall ends at the target. Stretching the last bar to close the gap without saying so is the error the exhibit exists to avoid.

P4. The cost of a hardcode (US). Cascade's FY2 pre-tax income is $134.6861 million. The assumption sheet says 25% and the formula says 21%. Compute the tax understatement, the net income overstatement in percent, and the effect on value per share given a correct value of $28.4316 and a hardcoded value of $28.5490. Then say what you tell the room.

Solution. Tax should be 134.6861 × 0.25 = 33.6715 and is 134.6861 × 0.21 = 28.2841, understated by $5.3874 million. Net income is 106.4020 against 101.0146, overstated by 5.3874 / 101.0146 = 5.3333%. Value per share moves by 28.5490 / 28.4316 − 1 = 0.4130%, or 11.74 cents. Tell the room the number, the effect and the fix date, in that order, and send the corrected file the same day. Explain that a 0.41% effect doesn't matter and you lose the mandate. The reader is pricing the process that produced the error, and the error's size hardly enters into it.

P5. The weighted value and its sensitivity (India). Bear ₹693.0779 at 0.25, base ₹1,303.9167 at 0.55, bull ₹1,829.5563 at 0.20. Compute the probability-weighted value. Then recompute with the bear at 0.40 and the base at 0.40, and state the change in percent.

Solution. 0.25(693.0779) + 0.55(1,303.9167) + 0.20(1,829.5563) = 173.2695 + 717.1542 + 365.9113 = ₹1,256.3349. Reweighted: 0.40(693.0779) + 0.40(1,303.9167) + 0.20(1,829.5563) = ₹1,164.7091, a change of −₹91.6258, which is −7.2931%. Show both, and a reader who thinks your bear case is more likely than you do can see straight away whether that disagreement changes the recommendation. Here it doesn't, since ₹1,164.71 is still 15.4% above the traded price.

P6. Reading a tornado (US). The five swings are $7.936, $5.304, $4.972, $4.316 and $2.055. Compute the total, the top two drivers' share, and say where you spend the next hour of work.

Solution. The total is $24.5828. The top two are the discount rate and capital intensity, whose swings total $13.2396, and 13.2396 / 24.5828 = 0.538569, or 53.9%. The next hour goes to the discount-rate build. A 20 basis point cut adds $1.0078 a share against $0.5153 for a full extra point of annual revenue growth, a ratio of 1.96, and revenue growth carries only 8.4% of the total swing. Analysts default to the revenue build because it's the part they enjoy. The tornado is the exhibit that argues against that instinct.

P7. Share against contribution (India). Segment revenue moved from 1,428, 924, 672 and 336 to 1,449, 1,323, 1,050 and 378 (₹ crore). Compute each segment's share of the two-year revenue increase, and choose the exhibit for a thesis page claiming a mix shift.

Solution. The increase is 840 crore: 21, 399, 378 and 42 by segment, which check to the total. Shares of the increase are 2.5%, 47.5%, 45.0% and 5.0%, so pharma intermediates and performance chemicals supplied 92.5%. The exhibit is a sorted horizontal bar of contribution to growth, because the message is a comparison across categories with no intrinsic order. Put the two named segments in the message colour and the other two in grey. Two pies would show share and miss contribution, would ask the reader to compare eight angles, and would break the rule against pies above three parts.

P8. The peer line (US). The fitted relationship across five peers is EV/EBITDA = 8.599174 + 1.121784 × growth in percent, with an R-squared of 0.991817. Cascade grows at 7.0% and trades at 13.4 times against a peer mean of 16.16 times. Compute the fitted multiple and the residual, and say whether growth explains the discount.

Solution. Fitted is 8.599174 + 1.121784 × 7.0 = 16.451664 times, so the residual is 13.4 − 16.451664 = −3.051664 turns. Growth explains none of the discount, and in fact makes it worse. Cascade grows faster than the peer average of 6.74%, so the relationship puts it +0.291664 turns above the mean. The raw discount of 2.76 turns therefore understates the unexplained gap. State the sample size beside the R-squared, because 0.99 across five points describes a fitted line and proves no law.

P9. The moving crossing point (India). With the left axis fixed at 28 to 33, the two plotted lines cross at quarter index 1.3608 when the right axis runs 95 to 120, at 1.5207 when it runs 90 to 130, and at 1.9243 when it runs 80 to 160. Compute the movement between the widest and narrowest choices and name the two acceptable repairs.

Solution. The crossing moves 1.9243 − 1.3608 = 0.5635 of a quarter, and 1.5207 − 1.3608 = 0.1598 between the first two. Each of the three axis ranges is defensible, so the crossing point is an authorial choice wearing the costume of a finding. The first repair is to index both series to 100 at the first period and plot them on one axis, which for these data ends at 103.2051 against 98.4000, a spread of 4.8051 index points. The second is small multiples: two stacked panels sharing an x-axis, with each y-axis labelled in its own units.

P10. Trimming a grid (US). The full five-by-five grid runs $21.2792 to $41.2877 against a base case of $28.4316. The middle three by three runs $24.4098 to $33.7902. Compute both ranges as percentages of the base case and the share of the full range the trim keeps, then write the footnote.

Solution. Full range 41.2877 − 21.2792 = 20.0085, which is 70.3741% of 28.4316. Trimmed range 33.7902 − 24.4098 = 9.3804, which is 32.9930% of the base and 9.3804 / 20.0085 = 46.8824% of the full range, achieved by removing 16 of 25 cells. The footnote names the exclusions in terms of the drivers. It says that discount rates below 8.9% sit under any defensible cost-of-capital build for this issuer, and terminal growth above 2.9% approaches long-run nominal economic growth. A footnote that justifies the trim by the tightness of the resulting answers has chosen the exhibit's message from its output.

P11. The source audit (India). Six front exhibits carry 47 distinct numbers, of which 31 have a source line. Compute the coverage and the number of lines to add. Then classify these four: the FY26 segment revenue of ₹1,323 crore; the 92.5% contribution share; the 11.7% discount rate; the traded price of ₹1,009.22.

Solution. Coverage is 31 / 47 = 65.9574%, leaving 16 numbers, or 34.0426%, without provenance. Segment revenue is reported, so its line names the annual report, the segment note and the year. The 92.5% is computed, so its line names the two reported inputs and the operation. The discount rate is estimated, so its line says so and names the basis, such as a cost-of-equity build with its inputs. The traded price is a market number and carries a timestamp, because it was true at a moment. Every unsourced number is either an estimate you haven't labelled or a defect, and the four-word fix for the first case costs less than the argument the second case starts.

P12. Ordering the tour (US). A bottom-up walkthrough takes 22.0 minutes and an output-first walkthrough takes 9.0. Compute the saving in minutes and percent, then write the three blocks with their contents for the Cascade model.

Solution. The saving is 13.0 minutes, which is 13.0 / 22.0 = 59.0909%. Block one, three minutes, the output: $28.43 a share, enterprise value $1,418.1 million, terminal share 83.3%, implied market price $45.59, and the three integrity checks. Block two, three minutes, the drivers that move it: the discount rate at 32.3% of the swing, capital intensity at 21.6% and the EBITDA margin at 20.2%, each located on the assumption sheet with its source and its plausible range. Block three, three minutes, mechanics: the assumption sheet, one driver row's formula, the property roll-forward and the cash sweep. The three blocks add up to the nine minutes.

Applied mini-project

Take a valuation you've already built, from the modelling phase or from your own work, and turn it into a research deck and a walkthrough. If you don't have one, the supplied Vindhya driver sets and the Cascade model are enough to complete every artefact.

Deliverable. An eleven-page front section, an indexed appendix, a one-page assumption sheet, a source register and a nine-minute tour script.

Artefact 1, the thesis page. Claim, market case, analyst case, falsifier, rating definition, source line. Test 1. The market case must return the traded price to within 0.5% when run through your model. Show the run. Test 2. Cover the exhibit. A reader must be able to restate the variant perception in one sentence containing a number.

Artefact 2, the valuation bridge. Price to target, one bar per driver. Test 3. The bars sum to the gap, or an interaction bar is drawn and labelled and the sum then ties. Show the residual to four decimal places. Test 4. The footnote states the order and the appendix carries the alternative construction.

Artefact 3, the scenario page. Three coherent cases with probabilities summing to 1. Test 5. Each case carries an observable trigger with a threshold and a date. Test 6. The page states the probability-weighted value, the reward-to-risk ratio and the probability at which the call fails.

Artefact 4, the risks page. Four columns: risk, observable, effect on value, known by. Test 7. Every magnitude in the effect column already appears elsewhere in the deck. No new quantities. Test 8. Any finding that argues against the recommendation appears here and stays out of the appendix.

Artefact 5, the assumption sheet. Every input, with value, type and source. Test 9. The three integrity checks pass and their residuals are printed on the output page. Test 10. A reader picking any number off any exhibit can find it on the sheet or trace it to a reported source in one step.

Artefact 6, the source register. One row per number on the front exhibits. Test 11. Coverage is at or above 95%. With 47 numbers that permits 2.35 unsourced, so in practice two.

Artefact 7, the tour script. Three blocks of three minutes. Test 12. Rehearsed once against a clock, the tour reaches the value per share inside the first ninety seconds.

The rubric runs to 24 points, with a pass at 19 and no zero permitted on any row.

Any artefact whose test fails scores zero on its row, however good the page looks. A beautiful deck whose bridge doesn't tie has failed at the only thing a research deck is for: being checkable.

Reading & resources

1. Barbara Minto, The Pyramid Principle. Paid, intermediate. The governing-thought idea and the habit of grouping supports under a claim. Read the sections on ordering ideas. The rest of the book is about written documents, and it belongs to the memo more than to the deck.
2. Gene Zelazny, Say It With Charts. Paid, beginner. The oldest and still the clearest treatment of matching a chart form to a message. The message-first sequence is the direct ancestor of the chart grammar you've been using.
3. Cole Nussbaumer Knaflic, Storytelling with Data. Paid, beginner. Decluttering, pre-attentive attributes and the practice of stating the message in the title. The chapter on removing chart junk is the fastest single improvement available to most analysts.
4. Stephen Few, Show Me the Numbers. Paid, intermediate. Table design, which matters more in research decks than anywhere else because the variant-perception exhibit and the assumption sheet are both tables.
5. Edward Tufte, The Visual Display of Quantitative Information. Paid, intermediate. The data-ink argument and small multiples, the technique that replaces every dual axis in a research deck.
6. Alberto Cairo, How Charts Lie. Paid, beginner. The dual-axis and truncated-axis chapters give the reader's-eye view of the crossing-point artefact demonstrated in the fifth worked example.
7. Aswath Damodaran's valuation writing and public spreadsheets. Free, advanced. The relevant part here is his treatment of what a terminal value assumption actually asserts, which is the argument the model output page has to survive.
8. SEBI (Research Analysts) Regulations, 2014, and the circulars issued under them. Free, advanced, India. Read the disclosure requirements in the original, and skip the summaries. Verify the current text against the regulator's own site, checked 8 September 2026.
9. Regulation Analyst Certification and FINRA Rule 2241. Free, advanced, United States. The certification language and the research-analyst conflict rules for member firms. Verify current text and applicability against the primary rulebooks, checked 8 September 2026.
10. CFA Institute's Standards of Practice Handbook, the sections on investment analysis, recommendations and actions. Free summary, paid full text, intermediate. Reasonable basis, record retention and the distinction between opinion and fact, which is the standard your sourcing work serves.
11. Any three sell-side initiation decks in your sector, read for structure rather than content. Free where public, beginner. Count the pages before the first exhibit, find the falsifier if there is one, and time how long it takes you to locate the ratings definition.
12. The FASB and Ind AS segment reporting standards. Free, advanced. Worth twenty minutes because the decomposition exhibit lives or dies on whether the reported segments mean what you think, and restated prior-year segments are the commonest cause of a contribution table that doesn't sum.

Teach it back & journal

Feynman prompt. Find someone who can read a spreadsheet and has never sat in a research meeting. In ten minutes, and using one traded price and one target, teach them what a variant perception is and why a valuation bridge can be drawn two different ways that both land on the same target. Do it without the words thesis, storyline or waterfall. You'll know the explanation worked when they ask, unprompted, which bar order you chose and why. Asking that means they've understood that the bars tell a story about the gap, and the story has an author.

Then teach the second half. Show them the three integrity checks and ask them to find the hardcoded tax rate in the model. Watch what happens to their questions after they find it. Almost everybody moves from asking about the company to asking about the model, and they don't move back. That shift is the reason the checks go on the output page. A footnote would bury them.

Journal prompt. Take the last piece of analysis you presented to anyone, in any format. Write down, in one sentence each, three things. What the audience already believed before you started. What you claimed that differed from it, stated with a number. What observable would have proved you wrong.

If you can't fill the second line with a number, you presented a description with no argument in it, and the audience's polite questions were the only feedback you were ever going to get. If you can fill the second line but not the third, you made a claim you had no intention of being held to. That's the harder habit to break, because it feels like confidence.

Then write the falsifier you would have used, with a threshold and a date, and put it in your calendar on that date. In three months you'll know something about your own judgment that no amount of deck craft can substitute for. And writing it down before you know the outcome is the only version of that lesson worth anything.