Outcomes

How Signal Outcomes Are Tracked

Every signal Fahali generates is tracked through a complete lifecycle — from detection to resolution to accuracy measurement. This outcome tracking is the foundation of honest accuracy reporting.

Signal Registration

When a signal is generated, the learning engine registers it in the spectra_signal_outcomes table. Each registration includes: the symbol, the engine that produced it, the alert type, the confidence score, a timestamp, and a forecast horizon. Signals are time-bound — they predict an outcome within a specific window (1h, 24h, or 48h), not indefinitely.

Outcome Resolution

After the forecast horizon elapses, the resolver checks whether the predicted outcome occurred. The result is one of three states:

Correct — the market moved in the predicted direction (or volatility expanded, or the crash occurred) within the horizon
Incorrect — the market did not move as predicted within the horizon
Expired stale — the signal could not be resolved (e.g., insufficient price data, symbol delisted)

Base-Rate-Adjusted Scoring

Raw win rates are misleading. A signal that predicts "down" in a market falling 70% of the time will have roughly 70% accuracy by chance alone. Fahali adjusts for this by measuring lift — the improvement over the base rate. If the market has a 55% directional bias and the signal achieves 75% accuracy, the lift is +20 percentage points. This is the figure that matters.

4-Axis Scorecard

Not all signals predict the same thing, so a single accuracy number is not meaningful. Fahali segments outcomes by prediction type:

Direction — up/down/neutral price prediction
Magnitude — the move exceeded a minimum size threshold
Volatility — volatility regime change (expansion/contraction)
Crash catch — extreme event detection

Each axis is scored independently. An engine may have high precision on crash catch and low precision on direction — that is expected. The scorecard makes this transparent.

Per-Engine Accuracy

Every 15 minutes, the system snapshots per-engine accuracy into the spectra_engine_accuracy table. This captures win rate, total signal count, and confidence calibration for each engine. These snapshots accumulate over time, building an honest track record.

Case Studies

Signals with verified outcomes are available as case studies at /insights. Each case study shows the original signal, its confidence score, the forecast horizon, and the actual outcome. We publish both successes and failures — the goal is transparency, not a highlight reel.

Retention and Transparency

Signal outcome data is retained for 90 days. Accuracy snapshots are retained for 30 days. This means the publicly reported track record always reflects recent performance, not cherry-picked historical periods. For detailed methodology, see /accuracy.