1. What Is Signal Half-Life?

Consider a prediction signal s_t (e.g. ML-predicted excess return). We bucket observations by signal strength. For a given bucket b, we define the cumulative forward return over horizon h.

As we extend h, the incremental edge usually decays. The half-life is the horizon at which cumulative return reaches approximately 50% of maximum, or where the marginal information ratio falls below some threshold.

2. Measuring Decay Curves

Practical method:

1. Sort daily observations by signal value into quantiles (e.g. top 10%)
2. For each entry date t, track realised returns for h = 1,...,H days
3. Compute average cumulative return for each h
4. Plot decay curves: E[R_{0→h}^(top)] vs h

You typically see strong positive edge for small h, flattening or reversal beyond a certain horizon.

3. Half-Life for ML vs Simple Models

Research use-case: compare decay curves for:

• ML models (e.g. XGBoost, LSTM)
• Simple models (e.g. 12-month momentum, value)

Questions: Are ML signals shorter-lived or longer-lived? Do ML signals mostly capture microstructure/short-horizon effects? Does model complexity correlate with shorter half-life?

4. Stability Across Regimes and Universes

Half-life is not constant:

• in high-vol regimes, edges may compress (shorter half-life)
• in calm regimes, edges may extend
• in small caps, microstructure edges may be strong but short-lived
• in large caps, slower fundamental edges may dominate

Thus we measure half-life by regime, by asset bucket, and by horizon type.

5. Decay and Portfolio Construction

Knowing half-life helps:

• choose optimal holding periods
• avoid holding positions beyond their edge horizon
• design staggered rebalancing policies
• combine multiple signals with different half-lives

Example: ML microstructure signals (half-life 1–2 days) → trade with high turnover, tight risk. Medium-term factor signals (half-life 20–60 days) → trade with lower turnover, larger sizing.

6. How volarixs Measures Signal Decay

Because volarixs stores full prediction histories, tracks realised returns, and stores regime labels, we can compute decay curves and half-life metrics automatically:

• per experiment, per model, per horizon
• per asset bucket and per regime

Outputs:

• decay plots in the UI
• half-life estimates stored in run metadata
• filters in leaderboards: e.g. "show models with half-life between 2 and 5 days"

This shifts the focus from "Is this model predictive?" to "How long does this predictive edge persist, for which assets, and under which regimes?" Which is the question that actually matters for capital deployment.