Tools

What does a tiny edge actually look like?

A real quant signal has an information coefficient around 0.05 — it explains a quarter of one percent of the variance and shows up as a shapeless cloud. Yet it still sorts winners from losers, and with enough breadth that's a genuine edge. Drag the skill, the number of names, and the decay, and watch the moment noise turns into alpha.

What it shows: the information coefficient (IC) is the correlation between a signal's forecasts and the returns that follow. This explorer draws forecasts and outcomes with exactly the IC you set, so you can see three things a single number hides: an IC of 0.05 explains only IC² = 0.25% of the variance (the scatter looks like noise), yet sorting on the forecast still makes the top group beat the bottom by ≈ IC × 2.8 standard deviations — and whether that edge is even detectable depends on your breadth (t ≈ IC·√N). A third slider fades the edge with a half-life, the way real signals decay.

Everything is synthetic and illustrative — no real signal, no performance claim, not trading advice. The forecast/outcome pairs are a teaching device (a standardised bivariate-normal draw); the quintile payoff, hit-rate, and significance are the standard closed forms (validated against simulation before this was built), and the scatter is one random sample so its measured IC wobbles around the truth — which is the whole point about telling skill from luck. The concepts are real: the Fundamental Law of Active Management (Grinold 1989; Grinold & Kahn), IC significance and breadth (Clarke, de Silva & Thorley 2002), and alpha decay. Related tools: Information Coefficient Calculator · Signal Decay Calculator · Signal Combination Simulator · Backtest Overfitting Simulator · Cointegration & Pairs Trading Simulator · Correlation Heatmap · VPIN & the Volume Clock.

Dial in an edge

Set the true skill of the signal, how many names you rank each period, and how fast the edge fades. The scatter is one random sample of that edge; the payoff, significance, and decay update from the maths.

0 = no skill (pure noise) · 0.05 = a strong real-world signal · 0.30 = implausibly good

how many bets the edge is spread across — more breadth makes a small edge detectable

periods until the edge is cut in half — shorter means you must trade it sooner

re-draws the scatter from the same true edge — watch the measured IC jump around

A real, detectable edge — small but significant

An IC of 0.05 explains just 0.25% of the variance — the scatter looks like noise — yet sorting on the forecast makes the top quintile beat the bottom by 0.14σ, with a directional hit-rate of 51.6%. Across 2,000 names that edge is a 2.2σ result — statistically detectable. And it fades: with a 6-period half-life it decays into the noise floor after about 14 periods, so the whole game is harvesting it before then.

What this edge is worth

Variance explained

0.25%

R² = IC² — why it looks like noise

Top−bottom quintile

0.14σ

the edge you actually trade

Detectability

2.2σ

significant @95%

Hit-rate

51.6%

right direction (½ + asin·IC⁄π)

Measured this draw

0.07

vs true 0.05 — sampling noise

Breadth

2,000

names ranked

Half-life

6

periods to half the edge

Useful life

~14

periods until it's noise

An edge this small looks like noise

true edge: y = 0.05·xForecast (standardised) → · red = bottom 20% · teal = top 20%+3.6outcome

Each dot is one name: its forecast (x) against the outcome that followed (y). The teal line is the true edge (y = IC·x); the cloud scatters around it so widely that at a realistic IC you can barely see the tilt. This particular sample measured an IC of 0.07 — re-draw it and that number jumps around the true 0.05, by more when you rank fewer names.

…but it still sorts winners from losers

Q1-0.07Q2-0.03Q3+0.00Q4+0.03Q5+0.07higher avg outcome ↑Sort names into 5 groups by forecast → average outcome per group (σ)

Rank the names by forecast and split them into five equal groups; this is the average outcome in each. Even an invisible-looking edge produces a clean monotone staircase — the best-forecast quintile beats the worst by about IC × 2.8 = 0.14σ. That spread, repeated across many names and many periods, is the whole business. It's the same logic the Fundamental Law scales with breadth.

And it fades — so timing matters

noise floor IC≈0.01half-life — edge halved0.050Holding horizon (periods) → · the IC you can still harvest

A signal's edge doesn't last: here it halves every 6 periods (IC(h) = IC·2^(−h⁄half-life)), sliding toward the noise floor after about 14 periods. The shaded area is roughly the edge you can still harvest — rebalance near or inside the half-life and you keep most of it; hold too long and you're trading noise. Measure your own from data with the Signal Decay Calculator.

Reading this

Three numbers decide whether a signal is worth trading, and a single IC hides all of them: how big the edge is (tiny — and that's normal), how much breadth you have to make it detectable, and how fast it decays. A 0.05 edge that looks like nothing is a career; the skill is in proving it's real and harvesting it before it's gone. Bring your own forecasts to the IC Calculator to measure the real thing.

Embed this explorer on your site

Free to embed — please keep the short credit line below the widget.

<iframe src="https://microalphas.com/tools/signal-skill-explorer/" width="760" height="1520" frameborder="0" loading="lazy" title="Signal Skill Explorer"></iframe>
<p style="font:13px/1.4 sans-serif;margin:6px 0 0">Powered by the <a href="https://microalphas.com/tools/signal-skill-explorer/">Signal Skill Explorer</a> from <a href="https://microalphas.com/">Micro Alphas</a></p>