IC measured at a one-month forward horizon is not the whole story. Measure the same factor's IC against 3-month and 6-month compounded forward returns instead, and you learn how quickly its predictive power decays, information every real trading desk needs before choosing a rebalance frequency. Mechanically, decay happens because the information a signal captures gets absorbed into the price over time: a momentum score that correctly identifies this month's winners is describing a trend that other market participants also eventually notice and trade on, and a value score's cheapness gap tends to narrow precisely because cheap-relative-to-earnings stocks that stay cheap for a long stretch either re-rate upward (closing the gap the signal was measuring) or reveal that the market had a good reason to price them that way after all. Either way, the specific cross-sectional gap your factor score measured this month is not a fact that stays true indefinitely, it is a snapshot that new information, price adjustment, and the passage of time all steadily erode, which is exactly why an evaluation pipeline needs to measure decay explicitly rather than assume a one-month IC simply "keeps working" for as long as you hold the position. This module's compound_return(period_returns) (∏(1 + r) − 1 over a list of consecutive single-period returns) is the utility that builds a multi-month forward return out of the bundled panel's own monthly fwd_return column, so you can re-run information_coefficient at any horizon you like without needing new data.
Advertisement
How fast do signals decay, and what follows for rebalancing?
Real numbers: momentum's mean IC runs 0.1209 (1 month) → 0.0632 (3 months) → 0.0496 (6 months), a clean, fast decay, more than half its one-month edge gone by the third month. Value decays more slowly in relative terms: 0.1013 → 0.0327 → 0.0402 (the small 3-to-6-month uptick is itself informative, real decay curves are rarely perfectly monotonic once you are averaging over a modest 43–46 overlapping windows, and a single bump should not be over-read as a trend reversal). Quality: 0.0805 → 0.0395 → 0.0425, a similar shape. Low-volatility decays from nothing to nothing: −0.0012 → 0.0109 → −0.0011, noise at every horizon it is measured at, a signal with no edge at one month has no edge to decay at three or six, either.
The practical implication, and the direct link forward to QD1.03: a fast-decaying signal (momentum, here) only pays off if you rebalance often enough to capture its edge before it fades, but rebalancing often is exactly what drives up the turnover measured on the portfolio turnover page, which drives up the transaction-cost drag QD1.03 will charge against the strategy. A slower-decaying signal buys you more room to rebalance infrequently and cheaply. Decay and turnover are not two unrelated statistics you happen to compute back to back; they are the two halves of the same rebalance-frequency trade-off.