ACCURACY · VERIFIED OUTCOMES · 291M CORPUS

Every detection tracked to resolution.

Fahali publishes a live accuracy scorecard. Every signal is tracked from creation to outcome against subsequent price action. Correct and incorrect are measured on the same scale. There is no separate scorekeeping.

88%
Magnitude precision
vs 54% move-base · +34pp lift
80.5%
Crash detection precision
vs 42.6% down-base · +37pp lift
62%
Directional precision
vs 55% trend-base · +7pp lift
291M
Labeled outcomes
Growing by 2M/day

Scorecard: Direction, Magnitude, and Volatility

The scorecard measures detection performance across four axes. Directional engines predict market direction (up/down). Magnitude engines predict a move exceeding 0.5%. Volatility engines predict regime expansion. Crash detection predicts a rapid decline exceeding 2% within 72h.

On the low directional lift: The +7pp directional precision lift is lower than our other metrics because 6 of 10 directional engines were temporarily operating as directionless (emitting neutral signals) due to an undiscovered taxonomy mapping issue. This was corrected on June 5, 2026. Directional data prior to that date includes these suppressed signals. Expect lift to improve as the corrected engine taxonomy accumulates resolved outcomes.
EngineAxisPrecisionBase rateLiftSignals
VolatilityMagnitude93.2%54.1%+39.1pp18,412
MomentumDirection65.1%55.0%+10.1pp21,048
Dark PoolMagnitude86.3%54.1%+32.2pp14,672
Order FlowDirection64.8%55.0%+9.8pp19,204
Regime DetectionVolatility89.7%58.3%+31.4pp11,554
Crash PredictorCrash80.5%42.6%+37.9pp6,201
CorrelationVolatility87.2%58.3%+28.9pp9,887
Order ImbalanceDirection61.3%55.0%+6.3pp22,418
LiquidityVolatility85.0%58.3%+26.7pp12,338
SqueezeMagnitude84.1%54.1%+30.0pp8,109
FundingMagnitude82.6%54.1%+28.5pp5,217
Capital FlowDirection59.4%55.0%+4.4pp15,663
SkewVolatility83.8%58.3%+25.5pp7,441
Volume SpikeMagnitude91.5%54.1%+37.4pp16,892
ReversalMagnitude87.6%54.1%+33.5pp10,235
DivergenceMagnitude88.4%54.1%+34.3pp9,718
ConfluenceMagnitude93.8%54.1%+39.7pp7,016
StealthMagnitude85.5%54.1%+31.4pp11,906

How the scorecard works

Signal resolution process

Every alert created by a Fahali detection engine is assigned a unique ID and recorded in the outcomes ledger with its engine, symbol, timestamp, predicted direction (if directional), and confidence score. After the forecast horizon expires (1-72 hours depending on engine type), the outcome is resolved against actual price action. The outcome is then classified into one of four axes: direction (bullish/bearish correct/incorrect), magnitude (>0.5% move detected), volatility (regime expansion), or crash (>2% decline within 72h). Blended precision is the weighted average across all engines within each axis.

Base-rate comparison

The base rate is calculated from a 90-day rolling window of the same metric. For magnitude, it is the frequency of 1-hour windows where absolute return exceeds 0.5%. For direction, it is the frequency of the more common direction in 1-hour windows. For crash detection, it is the frequency of 72-hour windows where decline exceeds 2%. Base-rate lift = precision − base rate. A positive lift means the engine performs better than a naive "always guess the most common outcome" strategy — the only metric that separates genuine signal from market drift.

Cooldown and deduplication

To prevent the same signal from being counted multiple times in consecutive scans, Fahali enforces a 900-second (15 minute) cooldown per (symbol, alert_type) pair. During the cooldown window, duplicate alerts are suppressed at the write boundary. This ensures the scorecard reflects the accuracy of the detection, not the frequency of its activation. The deduplication layer operates independently of the detection engines — no engine is aware of another's cooldown state.

What the numbers do not tell you

The scorecard measures precision (how many of our alerts are correct), not recall (how many correct alerts we could have produced). This is intentional: in risk surveillance, false positives erode trust faster than missed opportunities. Fahali prioritizes precision over recall because an alerting system that cries wolf is tuned out. The 7-engine consensus layer and 900-second cooldown are both precision-maximizing design decisions.

If Fahali were optimized for recall, we could emit alerts on every scan cycle and catch every move. The scorecard would show 95%+ recall. But the false positive rate would destroy usefulness. The 80-88% blended precision you see reflects the deliberate design choice to only alert when the evidence is strong and confirmed across independent engines.

Update June 5, 2026: A taxonomy issue affecting 6 of 10 directional engines (they were emitting directionless signals) was corrected. The directional lift is expected to improve as the corrected engines accumulate resolved outcomes over the coming weeks. The magnitude and crash numbers were unaffected because those engines were already operating correctly.

Verify the numbers yourself.

14-day trial · No card required · Every signal visible in your dashboard