Quantitative Signal Decay: How Long Does a Factor Edge Last and What To Do When It Fades

Every systematic strategy is discovered through research. The act of discovering it — publishing it, trading it, telling others about it — begins to destroy it. Signal decay is not a theoretical concern reserved for academic debates about market efficiency. It is the central operational challenge of running a live quant book. The question is not whether your edge will decay. It is when, at what rate, and what you will do when the metrics confirm it.

McLean and Pontiff (2016) quantified this precisely for published equity factors: factors lose approximately 32% of their pre-publication return in the three years following publication. The mechanism is well understood — institutional capital flows into the factor, compressing the spread between the long and short legs. But publication-driven crowding is only one of three causes of decay, and conflating them leads to wrong management responses. The first step in preventing premature signal retirement is building the research pipeline correctly from the start — see our guide to the quantitative alpha research process.

The Problem: Every Edge Has a Half-Life

Three distinct mechanisms drive signal decay. They require different responses, so diagnosing which is active matters more than knowing that decay is occurring.

Factor Crowding

As more capital pursues the same signal, the price impact of trading it compresses gross returns. When a momentum signal was first systematically exploited in the 1990s, buying the top-quintile winners and shorting the bottom-quintile losers moved the prices of those securities modestly relative to the available edge. By 2010, institutional systematic funds ran momentum signals across nearly every liquid equity market globally. The gross IC that a pure momentum signal generated in the 1990s could not survive the execution costs of the 2010s at scale. The McLean-Pontiff estimate — 32% return compression in three years post-publication — is the lower bound. Unpublished proprietary signals that become industry standard through informal channel-sharing often decay faster.

Regime Change

Structural breaks in market microstructure alter the data-generating process underlying a signal. Decimalization in 2001 destroyed the microstructure edge that many short-term reversal strategies exploited in the pre-decimal era. HFT proliferation between 2005 and 2010 made intraday mean-reversion signals uneconomic for any fund that was not itself operating at sub-millisecond latency. The zero-rate era from 2009 to 2022 suppressed value factor performance by sustaining growth stock valuations at multiples that would have reverted much faster in a higher-rate environment. Post-COVID vol regime starting in 2022 compressed carry strategy performance as macro volatility structurally increased. In each case, the underlying signal was not wrong — the economic regime that made it profitable changed. This is the category of decay that recovers when the regime reverts, which makes it the hardest to distinguish from the other two.

Overfitting Decay

Signals discovered through data mining degrade because the edge was noise, not a true inefficiency. The IC in the backtest was never real — it was the artifact of mining a sufficient number of parameter combinations in a finite historical dataset. A signal with no economic mechanism to explain why it should work, selected from a broad search of indicator combinations, is almost certainly a noise discovery. The Deflated Sharpe Ratio framework (Bailey and López de Prado, 2014) quantifies how much of a backtest's apparent Sharpe is explained by multiple testing bias. For a factor that was selected from 50 trial combinations, a reported backtest Sharpe of 0.9 can imply a DSR-adjusted Sharpe of 0.3 — below any viable deployment threshold before accounting for live execution costs. Overfitting decay produces an IC that was always going to be zero or negative out-of-sample. There is no recalibration response that fixes it. The correct action is retirement and reallocation of research bandwidth.

Measuring Decay: The Metrics That Matter

Four metrics together provide a complete picture of live signal health. Any one of them alone is insufficient. The decay diagnosis requires the full panel.

(a) Rolling Information Coefficient

The IC is the Pearson correlation between the factor score at lag k and realized return:

For a cross-sectional equity factor, the standard lag is 1–5 trading days for short-term signals and 21 days for monthly rebalancing signals. Compute IC on a rolling 20-day and 60-day window for every live signal simultaneously.

Operational thresholds: a healthy signal maintains IC greater than 0.03 consistently gross of costs. IC below 0.02 on a rolling 60-day window for three or more consecutive months is a warning signal — the edge is approaching the point where it cannot survive realistic transaction costs. An IC sign flip — negative IC for 30 or more consecutive trading days — is a red flag. A signal generating negative IC is actively losing money on its directional bet, which means the factor relationship has reversed and the position sizing engine is taking the wrong side.

(b) Out-of-Sample Sharpe Ratio Degradation

The in-sample (IS) Sharpe ratio from the research backtest is your baseline, but it must be adjusted using the Deflated Sharpe Ratio framework before setting the out-of-sample (OOS) threshold. If the IS Sharpe was selected from a broad parameter search, the true expected Sharpe is materially lower. The OOS threshold should be calibrated against the DSR-adjusted IS Sharpe, not the raw observed backtest number.

Warning threshold: OOS Sharpe below 50% of the DSR-adjusted IS Sharpe for six or more months. This is a sign that either crowding or regime change has meaningfully compressed the live edge. Red flag: OOS Sharpe below zero for trailing 12 months. A strategy that is generating negative risk-adjusted returns over a full year is not in a temporary drawdown — it is in structural decay.

(c) Factor Crowding Index

When many funds are long the same high-momentum names, bid-ask spreads widen, price impact increases, and the signal self-destructs on execution. The crowding signal is not visible in IC alone because IC is gross of execution costs — a signal can show healthy gross IC while being completely consumed by widened transaction costs.

Two practical crowding proxies. First: compute the average correlation between the factor portfolio (the basket of longs and shorts generated by the signal) and known systematic factor benchmarks — AQR-style momentum factor portfolios, Cliff Asness's published factor returns, or major systematic fund factor tilts available in regulatory filings. High correlation to publicly available factor portfolios means many other funds are in essentially the same positions. Second: run factor-adjusted return attribution against publicly available ETFs. If a momentum signal's gross alpha can be explained at R² greater than 40% by the MTUM ETF, the edge is eroding into beta — you are paying active management fees and transaction costs for a position that could be achieved more cheaply via a passive vehicle.

(d) Transaction Cost Absorption Rate

Monitor the fraction of gross alpha consumed by transaction costs separately from portfolio-level cost analysis. A signal with high turnover and a tilt toward less-liquid instruments will show a very different cost absorption rate than a low-turnover value signal running in large-cap equities. The signal-specific capacity constraint — not the portfolio-level average — determines whether the live edge is viable.

Healthy: transaction costs below 30% of gross alpha. Warning: costs above 50% of gross alpha — the gross IC must be substantially above the warning threshold to generate acceptable net alpha at this cost level. Red flag: net alpha at or below zero — the signal is fully crowded or decayed in execution terms even if the gross IC is still technically positive.

The Decay Timeline: What the Research Says

Factor half-lives are not uniform. They vary systematically by factor type, liquidity tier, and the speed at which institutional capital can replicate the strategy. Understanding the expected half-life before a signal goes live sets the right monitoring cadence and research reinvestment timeline.

Value and Quality: Long Half-Lives (5–10 Years)

Value factors (book-to-market, earnings yield, cash flow yield) and quality factors (profitability, earnings stability, low leverage) reflect genuine risk premia and structural behavioral biases that are slow to arbitrage away. The mechanism — mean reversion in valuation multiples and fundamental earnings convergence — operates over multi-year horizons. Institutional capital cannot compress a multi-year mean reversion trade in the same way it can compress a short-term momentum signal. For a full treatment of the factor zoo survival analysis and which value and quality factors have genuine long-run persistence, see factor investing for hedge funds.

Value factors are not immune to long drawdowns — the decade-long underperformance from 2009 to 2020 was a regime-driven episode, not crowding-driven decay. That distinction matters for the retirement decision. A 10-year value drawdown during a zero-rate regime is not evidence that the factor is dead — it is evidence that the regime is suppressing the mechanism. Value recovered sharply in 2022 when rates normalized. The correct management response was Hold, not Retire.

Short-Term Momentum: Medium Half-Life (2–5 Years)

Cross-sectional equity momentum (12-month return minus most recent month) has one of the strongest published Sharpe ratios in the academic literature and has proven more durable than most factors post-publication. But it is subject to regime change in ways that value is not. The 2009 momentum crash — when a sharp reversal from the GFC lows unwound the momentum trade catastrophically in a single month — and the 2022 compression as macro volatility increased illustrate the cyclical fragility of medium-term momentum. The factor is not dead after these episodes; it requires regime-aware position sizing and explicit crash risk management to harvest it at institutional scale consistently.

Short-Term Reversal: Fast Decay (<12 Months Post-Publication)

Short-term reversal (1-week or 1-month reversion) is the most vulnerable factor category to HFT competition. The mechanism — transient liquidity imbalances reverting as market makers absorb order flow — operates on intraday to weekly timescales, and HFT strategies can capture the reversion in milliseconds. For a systematic fund operating at daily frequency, HFT has compressed the accessible reversal premium to near zero in U.S. large-cap equities. Residual alpha in reversal factors survives only in less-competed segments: small-cap stocks with wide bid-ask spreads, off-hours sessions, or international markets with lower HFT penetration. For the full treatment of how HFT competition has restructured the stat arb and reversal factor landscape, see statistical arbitrage strategies for hedge funds.

Machine Learning Alpha Signals: Fastest Decay (6–18 Months)

Alternative data signals — satellite imagery for retail foot traffic, NLP-derived earnings call sentiment, credit card transaction data, web scraping signals — are the fastest-decaying category in institutional practice. The institutional adoption cycle for new alt data sources has compressed to under 18 months. A proprietary signal built on a new alt data source in 2022 was being replicated by competing funds by late 2023. The problem is structural: alternative data vendors sell to multiple institutional clients simultaneously, making any signal built on a particular data source a shared signal within months of its commercialization. NLP models trained on earnings call transcripts, which required months of proprietary development in 2018, are now available as off-the-shelf APIs. For the full framework on building and deploying ML alpha signals with appropriate decay expectations built in, see machine learning in quantitative finance.

The meta-finding: the faster the publication-to-institutional-adoption cycle, the faster the decay. In 2010, a published factor had five or more years of live alpha before being fully arbitraged. In 2026, the equivalent timeframe is 18–24 months for most systematic equity and macro signals. The research reinvestment cycle has to match the decay cycle — you cannot build a durable quant book on signals that decay in 18 months if your research pipeline takes two years to generate a new deployable signal.

The Decision Framework: Retire, Recalibrate, or Hold

When the metrics flag decay, the management decision is threefold: retire the signal and reallocate the risk budget, recalibrate the signal to recover the edge, or hold and monitor pending additional evidence. Each has specific trigger conditions. Getting the wrong answer costs either capital (holding a decayed signal) or research opportunity cost (retiring a temporarily impaired signal that was recoverable).

Retire When

IC has been negative or zero for six or more consecutive months. OOS Sharpe has been below zero for trailing 12 months. The factor crowding index shows saturation at above the 80th percentile of historical crowding levels. Factor-adjusted attribution shows the signal's alpha is fully explained by a cheap public-factor ETF — the edge has become beta. There is no plausible economic mechanism explaining a recovery — the decay is a structural break in market microstructure, not a cyclical regime effect.

When retiring a signal, the priority is redeploying the risk budget immediately rather than letting it sit unallocated. A retired signal represents a slice of the portfolio's risk budget that needs a replacement with better forward-looking IC. For the portfolio-level mechanics of reallocating risk budget after a signal retirement, see quantitative portfolio construction and risk budgeting.

Recalibrate When

IC degradation is regime-specific — the factor worked in a low-vol regime but breaks in a high-vol regime, or the factor worked in a low-rate environment but does not in a high-rate environment. Transaction cost absorption above 50% but the capacity constraint is tractable — reducing position size, extending holding period, or tilting toward more liquid instruments recovers meaningful net alpha without requiring signal replacement. The signal is crowded in the most obvious segment (large-cap U.S. equity, the most-covered instruments) but alpha persists in less-competed sub-segments: small-cap EM, less-liquid factor variants, markets with lower institutional penetration. The underlying economic mechanism is still valid but the feature engineering needs updating — the NLP model needs a new fine-tune, the alt data provider changed their methodology, the signal parameters need rolling reoptimization.

Hold When

Short-term IC decline is regime-driven — during a VIX spike, factors that depend on normal-regime cross-sectional dispersion temporarily show degraded IC as correlated selling overwhelms alpha signals. For a full treatment of which signals survive high-vol regimes and which require pausing, see systematic trading in a high-volatility regime. Trailing 12-month Sharpe is below the IS threshold but trailing 36-month Sharpe is still acceptable and consistent with the IS backtest — one bad year does not constitute structural decay, especially for a factor with a documented 5–10 year half-life. The confidence interval on the IC trend is wide — small signal count, high signal volatility — and there is insufficient statistical evidence to distinguish retire from hold.

The 2×2 Decision Matrix

A simple framework for rapid triage across a large signal library. Two dimensions: IC trend (improving vs. declining over the trailing 60 days) and crowding level (below vs. above the 60th percentile of historical crowding). Four quadrants:

The bottom-right quadrant — IC declining and crowding high — is the unambiguous retirement trigger. Both primary decay causes (factor crowding compressing alpha and structural IC degradation) are active simultaneously. Waiting for confirmation in this quadrant costs capital. The top-right and bottom-left quadrants require investigation before acting; the top-left is a monitoring case only.

Building a Signal Lifecycle Management System in Production

The metrics and decision framework above are only useful if they are systematized. Ad hoc decay monitoring — a quant researcher pulling IC numbers when a signal has a bad quarter — is not signal lifecycle management. It is reactive damage control, applied after the signal has already cost capital. The governance requirement is a structured process with defined cadence, escalation thresholds, and documented decisions.

1. Daily IC report. At end of day, compute rolling 20-day IC and rolling 60-day IC for every live signal in the production book. This is not a weekly or monthly report — IC can cross warning thresholds in days during a sharp regime shift, and the daily cadence gives the research team the earliest possible visibility into deterioration. Flag any signal where rolling 60-day IC drops below 0.02, and separately flag any signal where IC has been negative for 20 or more consecutive trading days. The daily IC report is the primary monitoring tool; the other three reports investigate signals it flags.

2. Weekly OOS Sharpe tracker. Each week, compare every signal's year-to-date and trailing 12-month OOS Sharpe against its DSR-adjusted IS Sharpe benchmark. Compute the DSR-adjusted probability that the remaining edge is real versus noise — a signal that has been live for 12 months with an OOS Sharpe of 0.25 against a DSR-adjusted IS Sharpe of 0.6 has provided some evidence of live edge but not enough to be confident the backtest was predictive. For a signal that has been live for 36 months and has sustained OOS Sharpe of 0.5, the evidence is substantially more compelling. The weekly tracker forces that probabilistic accounting into the process rather than letting it remain implicit.

3. Monthly crowding audit. Run factor-adjusted return attribution for every live signal against the benchmark ETF and factor portfolio set. Flag any signal where the ETF or public-factor explained R² exceeds 40% — this is the threshold at which the signal is approaching commodity-factor territory. Also compute the transaction cost absorption rate for each signal and flag any signal where costs have exceeded 50% of gross alpha for two or more consecutive months. The crowding audit is slower-moving than the IC report but captures the most common silent decay path: a signal whose gross IC is still acceptable but whose net alpha has been quietly consumed by widening execution costs as crowding increases.

4. Quarterly lifecycle review. A structured review of every signal that has active warning flags — from the daily IC report, weekly OOS tracker, or monthly crowding audit — conducted by the senior PM or quant research committee. Each flagged signal receives a documented decision: retire (with explicit plan for risk budget reallocation), recalibrate (with specific engineering changes and timeline), or hold (with documented rationale and next review trigger). The quarterly review is a governance checkpoint, not a research session. The research work should be done before the review, so the committee is evaluating recommendations, not generating them in the room.

The governance point is this: funds that retire decayed signals early and redeploy research bandwidth to new edge discovery consistently outperform those that hold legacy signals “waiting for the regime to turn.” The waiting-for-regime rationalization is almost always wrong in cases of genuine crowding decay (the regime does not rescue a signal that has been competed away by institutional capital) and sometimes right in cases of cyclical regime change (value factor, 2009–2022). The discipline of formal lifecycle management forces the distinction. Without it, every decayed signal gets classified as regime-driven because regime-driven is the rationalization that does not require admitting a research error.

A practical signal governance benchmark: a fund running 20–30 live signals in a systematic equity book should expect to retire 3–5 signals per year in normal conditions and redeploy those risk budget slices into new signal research. If the retirement rate is zero — no signal has been retired in three years — the lifecycle management system is not functioning. Every factor eventually decays. A retirement rate of zero means signals are being held past their economically useful life, and the research pipeline is not being incentivized to replace them.

The Half-Life Is Shortening — Manage It Explicitly

Signal decay is not a crisis. It is the normal operating condition of a live systematic book. Every edge has a half-life. The half-life of most factors has shortened materially over the last 15 years as institutional adoption cycles accelerated. The quant book that treats signal lifecycle management as an operational process — with daily monitoring, formal quarterly reviews, and an explicit governance framework for retire/recalibrate/hold decisions — will consistently outperform the one that responds to decay reactively after a bad quarter.

The three causes of decay require different responses. Crowding decay is irreversible once capital has fully compressed the edge; retirement is usually correct. Regime-driven decay is cyclical; hold with monitoring and regime-aware de-risking is usually correct. Overfitting decay means the IC was never real; retirement is always correct. Getting the diagnosis right — using the four-metric panel together rather than any single metric in isolation — is the core research skill in running a live quant book. It is at least as important as the skill of discovering the signal in the first place.