Case study / Alpha Engine

6,788 signals. No measured edge.

The public Alpha Engine project replayed its directional calls against subsequent price moves. Its July 2026 evaluation found no measurable advantage over the direction-matched base rate.

Reported 27 Jul 2026 · Full findings ↗

The measured result
+0.0 pp

Directional edge over the matched base rate

6,788 scored signals across 7 assets, evaluated over the next 10 bars. This is a finding about this engine and this test, not about every possible trading signal.

Directional edge versus each asset's direction-matched base rate · percentage points
  • BTC -1.7 pp
  • ETH -0.7 pp
  • SOL -0.3 pp
  • AAPL +2.0 pp
  • MSFT -0.5 pp
  • GOOGL +0.1 pp
  • NVDA +0.7 pp

Reported in the July 2026 findings ↗. The spread is noise around zero; it does not establish an edge.

01Method

What was tested.

For each historical bar, the project replayed the same signal_at() path used by its scan. It scored each bullish or bearish call against the asset's move over the next 10 bars. The seven assets span crypto and US equities, with 2.7 to 5 years of history depending on the asset.

The baseline matters: if an asset rose in 56% of the test windows, a bullish call had to beat 56%, rather than a generic 50%. The aggregate result was +0.0 percentage points of directional edge over that matched baseline.

Inspect commands and per-asset results
02Corrections

The false positives were instructive.

Survivors looked better. Scoring only calls that had not hit a stop produced an apparent positive edge. That subset was selected by the outcome, so the comparison was invalid.
Half right was the wrong bar. Several calls looked acceptable against 50%. They did not beat the underlying asset's tendency to move in the called direction.
One correction was itself corrected. An early analysis treated stop hits as full losses and concluded that stops destroyed value. Return-based measurement did not support that conclusion. The project keeps the correction visible in its findings.
03Limits

What this result does not say.

It does not show that no strategy can work. It evaluates this engine's directional calls, these assets, this history, and a 10-bar horizon. The asset-level values range from −1.7 to +2.0 percentage points; the project interprets that spread as noise around zero.

The useful output is the measurement process and the record of ideas that failed under scrutiny. The full findings include the tables, commands, and later investigations. The repository contains the code.

For a smaller illustration of selection bias, read our reproducible Sharpe simulation.