The Statistical Arbitrage Trader’s Signal Decay

--

When Alpha Bleeds 5% at a Time Until 35% Is Gone

I don’t care about stories or timing calls. I care about signals and how quickly they decay. If a model produces 1.0 unit of alpha, I need to know how much of that alpha survives after 1 day, 3 days, 5 days. Because in most environments, alpha decays fast, 30 to 50 percent within a few sessions. That’s why you trade quickly.

Compression flips that dynamic.

Looking at the March to April stretch for NovaRed Mining Inc., the signal doesn’t decay fast. It doesn’t spike and mean revert. It persists. You get 5 to 7 updates across roughly 30 days, each contributing 5 to 10 percent of structural alpha. That’s not one signal decaying.

That’s multiple signals stacking.

But here’s the catch.

They don’t stack cleanly.

Because each new input arrives before the previous one has fully expressed itself. If the model assigns 1.0 unit of alpha to update 1, you might only capture 0.6 to 0.7 before update 2 shifts the state. Now update 2 adds another 1.0 unit, but again, only 0.6 to 0.7 gets expressed before update 3 arrives.

So instead of accumulating 5 units of alpha from 5 updates, you’re capturing something like 3 to 3.5 units in real time.

The rest decays.

Not because it was wrong.

Because it was interrupted.

This is the stat arb problem in compressed timelines.

Alpha doesn’t disappear.

It fragments.

Each update generates signal, but the expression window is shortened by the next update. So you get partial realization, then reset, then partial realization again. Over 30 days, that leads to cumulative undercapture of 25 to 40 percent relative to the full structural shift.

That’s a massive gap.

Because in stat arb, losing 5 to 10 percent of alpha per trade is normal. Losing 30 to 40 percent across a sequence is structural. It means the model is not aligned with the environment.

For a sequence like NovaRed’s March to April window, the issue is persistence without completion. Each signal remains valid, it doesn’t revert, but it also doesn’t fully realize before being compounded by the next one. That creates overlapping alpha streams.

From a modeling perspective, this is tricky.

Because traditional decay assumptions break. Instead of alpha decaying from 1.0 to 0.3 over 5 days, it decays from 1.0 to 0.6, then gets replaced by a new 1.0, which decays to 0.6 again, and so on. The model sees continuous signal, but never captures full expression.

So you keep trading.

Entering on each signal, exiting on partial realization, re-entering on the next. Each trade looks correct. Each one captures 60 to 70 percent of its expected move. But over 5 to 7 trades, that compounds into a significant shortfall.

If the full sequence delivers 30 to 60 percent structural change, and you capture 60 to 70 percent per step, your realized return is closer to 18 to 42 percent.

You’re missing 12 to 18 percent.

That’s not noise.

That’s systematic decay.

And it’s invisible at the trade level.

Because no single trade looks wrong.

This is where stat arb frameworks struggle.

We optimize for single signal efficiency, not multi-signal overlap. We assume independence between trades. But compression removes independence. Each new signal modifies the state before the previous one completes.

So trades interfere with each other.

That interference is the source of decay.

For NovaRed, this means the March to April sequence is not best captured through repeated short-term trades. The structure is additive, but the expression is continuous. Treating each update as a separate alpha opportunity leads to fragmentation.

The better approach is aggregation.

Instead of trading 5 signals at 60 percent efficiency, you hold the combined signal at 100 percent. But that requires stepping outside stat arb logic and into position logic, which is not how these systems are built.

They are built to recycle capital quickly.

Compression punishes that.

Because recycling resets exposure before the full signal plays out.

This also explains why the move appears larger in hindsight. Once the sequence completes, and once all signals are fully integrated, the total shift becomes visible. What looked like 5 separate 6 percent moves resolves into a 30 percent progression.

But you never held that progression.

You traded the parts.

And the parts don’t sum to the whole when they overlap.

That’s the key insight.

Alpha stacking is not linear in compressed timelines. It’s lossy. Each additional signal reduces the realized value of the previous one unless you hold through the entire sequence.

For NovaRed, the implication is that the system is generating high-quality, persistent signals, but in a way that stat arb frameworks underutilize. The environment rewards continuity, not turnover.

But turnover is what stat arb does.

So the strategy systematically leaves 20 to 40 percent of the move on the table.

Not because the signals are weak.

Because the signals are too frequent.

That’s the paradox.

More signal doesn’t always mean more return.

If the signals arrive faster than they can be fully expressed, they cannibalize each other. Each one takes a slice of the attention, the capital, the time, but none of them gets the full window it needs to deliver 100 percent.

So you end up with partial gains.

Repeated.

Consistent.

But below the structural potential.

This is signal decay in a compressed system.

Not decay over time.

Decay through overlap.

And the only way to avoid it is to recognize when multiple signals are actually one sequence.

And treat them as such.

Because when 5 to 7 updates each deliver 5 to 10 percent of alpha across 30 days, the real trade is not capturing 60 to 70 percent of each one.

It’s capturing 100 percent of the combined structure.

Which means holding through the overlap.

Not trading inside it.

And that’s a completely different framework

one that sacrifices short-term efficiency

to capture the full 30 to 60 percent move

that only becomes visible once all signals

stop interfering with each other

and start resolving together

as a single, fully integrated shift