alphafade

Is my trading signal dying, and if so, how fast and why?

Most tools measure how a signal's predictive power fades over the days after each trade. alphafade measures something different: how a signal's edge shrinks across calendar time (months and years), and whether that shrinkage lines up with crowding, meaning more money chasing the same pattern. It fits a decay curve and reports an honest half-life with a bootstrap confidence interval, or says "no detectable decay" when the data can't tell. It also tests for structural breaks, measures the post-publication drop McLean & Pontiff (2016) made famous, and computes a Lou & Polk comomentum crowding score. Every t-statistic is autocorrelation-robust (Newey-West), and nothing is dropped, resampled, or shifted behind your back.

Install

pip install alphafade             # core: numpy, pandas, scipy
pip install "alphafade[plot]"     # + matplotlib for report.plot()

Python 3.11 or newer. With uv: uv add "alphafade[plot]".

Quickstart

Is momentum dying? This downloads Ken French's momentum factor (a few KB, cached afterwards) and runs every analysis:

import alphafade as af

umd = af.datasets.load_momentum("M").loc["1963-07":]  # monthly momentum returns
report = af.analyze(
    umd,
    sample_end="1989-12-31",        # Jegadeesh & Titman's sample ended here
    publication_date="1993-03-01",  # ...and their paper came out here
    rng=42,                         # reproducible bootstrap
)
print(report.summary())
report.plot()                       # needs alphafade[plot]

Part of the output (data through August 2026):

Verdict: No detectable decay in the strategy's average return so far.
...
  After publication (1993-03 to 2026-08, 402 obs): average 0.003823 per period (4.59% a year,
  Sharpe 0.28, t = 1.60), 53% lower than in-sample (t of the change = -1.40).

Momentum earns about half as much since publication, but it's so volatile that 33 years of data can't rule out luck. alphafade tells you that instead of printing a confident-looking half-life.

With a signal instead of returns

If you have the signal itself (dates × assets) and asset returns, the per-date information coefficient (IC: the cross-sectional correlation between today's signal and the returns that follow) is usually the sharper thing to track:

import numpy as np, pandas as pd
import alphafade as af

rng = np.random.default_rng(0)
dates = pd.date_range("1980-01-31", periods=480, freq="ME")      # 40 years, 500 stocks
signal = pd.DataFrame(rng.standard_normal((480, 500)), index=dates)
edge = 0.10 * np.exp(-np.arange(480) / 12 / 5)                   # true half-life: 3.5 years
realized = 0.05 * (edge[:, None] * signal + rng.standard_normal((480, 500))).shift(1)

fwd = af.forward_returns(realized)   # row t = return earned AFTER t (no look-ahead)
ic = af.ic_series(signal, fwd)       # Spearman IC per date
fit = af.fit_decay(ic, rng=0)        # exponential decay + bootstrap CI
print(fit.summary())

Exponential decay fit on 479 observations (1980-01 to 2019-11).
The edge started at 0.07159 and is shrinking about 14.8% per year: half-life 4.3 years (95% CI 3.0 to 5.9).
The exponential model fits better than the linear one (AIC 22.6 lower).

The true half-life (3.5 years) is inside the interval. Across 12 random seeds of this setup the median estimate was 3.56 years and every interval contained the truth.

Beyond the basics

Four more tools, each answering a question a half-life alone can't:

• signal_lifetime turns a fit into "when does the edge reach a level I care about?"
• compare_signals ranks many strategies at once and corrects the p-values for testing many signals (otherwise some pure-noise signal will look like it is fading by luck).
• walk_forward_decay refits at each date using only the data available then, so you can see whether the half-life is stable or just a quirk of the full sample.
• ic_by_horizon measures fading across the forecast horizon (how many periods ahead the signal still predicts), which is a different question from fading across calendar time.

import numpy as np, pandas as pd
import alphafade as af

rng = np.random.default_rng(1)
dates = pd.date_range("1980-01-31", periods=480, freq="ME")
years = np.arange(480) / 12
fading = pd.Series(0.02 * np.exp(-years / 8) + rng.normal(0, 0.02, 480), index=dates)
noise = pd.Series(rng.normal(0.003, 0.02, 480), index=dates)

fit = af.fit_decay(fading, rng=0)
life = af.signal_lifetime(fit, fraction=0.5)     # when is the edge down to half its start?
print(life.summary())

ranked = af.compare_signals(pd.DataFrame({"fading": fading, "noise": noise}), rng=0)
print(ranked.table[["half_life_years", "p_adjusted", "decay_detected_adjusted"]].round(3))

walk = af.walk_forward_decay(fading, min_obs=120, step=60, rng=0)
print(walk.table[["n_obs", "decay_detected"]].tail(3))

The fitted exponential edge reaches 50% of its starting level about 3.6 years after the sample start (95% CI 2.3 to 5.7), around 1983-08. That has already happened. The interval holds the starting level fixed and only varies the decay rate.
        half_life_years  p_adjusted  decay_detected_adjusted
fading            3.554       0.000                     True
noise               NaN       0.129                    False
            n_obs  decay_detected
2009-12-31    360            True
2014-12-31    420            True
2019-12-31    480            True

The true half-life here is 5.5 years and the interval contains it. Multiple-testing correction lowers, but cannot remove, the chance that pure noise is flagged: with other seeds this same setup occasionally flags the noise series too. FadeReport.to_dict() and .to_json() export a report's headline numbers as plain data.

Public API

Errors are specific and say how to fix the input: InputError (a ValueError), AlignmentError, FrequencyError, InsufficientDataError, DownloadError, all subclasses of AlphaFadeError; warnings subclass AlphaFadeWarning. Anything lossy (dropping NaNs, thin cross-sections) emits a DataDroppedWarning with counts; unreliable fits emit FitWarning. Every function that uses randomness takes rng= (a seed or a numpy Generator).

How this differs from alphalens and quantstats

They're complements: use alphalens to build and vet a factor, quantstats to report on a portfolio, and alphafade to ask whether the edge is going away. alphafade isn't a backtester and doesn't build portfolios; you bring returns or a signal.

Limitations (read these)

• Decay is hard to measure. With a realistic signal (3,000 stocks, 60 years of monthly data, starting IC 0.10), single half-life estimates scatter by about ±6% (one standard deviation). For a volatile strategy's raw returns, decades of data often can't distinguish decay from noise, as the UMD example shows. Wide intervals are the honest answer, not a bug.
• One shift, not many. find_break looks for a single change in the average. Several regime changes, or a slow slide, show up as one "most likely" date. Use fit_decay to describe a gradual fade.
• Asymptotic p-values. sup-Wald p-values come from a simulated large-sample distribution (reproducible: scripts/make_supwald_table.py) and are floored at 0.0005. In simulations its false-alarm rate is close to 5% at a few hundred observations but can rise under strong autocorrelation. It gives up a little power to stay honest.
• Decay toward zero. The exponential model assumes the edge fades to 0, not to some permanent floor. A linear trend is always fitted alongside for comparison.
• Survivorship bias. If your stock universe contains only companies that exist today, the IC and crowding scores are biased (the losers that got delisted are missing). Use a point-in-time universe such as CRSP when you can.
• Frequencies are never guessed. Mixed daily/monthly data raises a FrequencyError; resample it yourself. Inputs must be wide (dates × assets) with a sorted DatetimeIndex.
• Weekly momentum. Ken French doesn't publish a weekly momentum file; compound the daily one.

Learn more

• docs/methodology.md: every formula and default, with references.
• examples/umd_momentum.py: the end-to-end momentum study above.
• CHANGELOG.md · CONTRIBUTING.md

References

McLean, R. D., & Pontiff, J. (2016). Does academic research destroy stock return predictability? Journal of Finance, 71(1), 5–32. · Lou, D., & Polk, C. (2022). Comomentum: Inferring arbitrage activity from return correlations. Review of Financial Studies, 35(7), 3272–3302. · Andrews, D. W. K. (1993). Tests for parameter instability and structural change with unknown change point. Econometrica, 61(4), 821–856. · Newey, W. K., & West, K. D. (1987). A simple, positive semi-definite, heteroskedasticity and autocorrelation consistent covariance matrix. Econometrica, 55(3), 703–708. · Jegadeesh, N., & Titman, S. (1993). Returns to buying winners and selling losers. Journal of Finance, 48(1), 65–91.

License

MIT © 2026 Jeevun Sandhu