Performance

Signal Performance

Every signal HyperLenz fires is scored against its realized market outcome after it expires. The table below is computed from all scored signals in the window — losers included. No cherry-picking, no survivorship bias. Win rate is the fraction of resolved signals that met their type's win condition — a move in the predicted direction beyond the flat deadband for directional signals, realized spread capture for market-making and mean-reversion signals. Undetermined signals are reported separately and excluded from the rate.

Row-based performance (v1)

Each row of signal_outcomes is scored as an independent result over the default 7-day window. A signal often re-fires on the same market while one move plays out, so this table counts correlated re-fires as separate results, which overstates how much independent evidence there is. The episode-aware view below collapses them.

Undetermined — a signal could not be scored: snapshots at fire time or at the scoring horizon were missing or stale, funding samples were insufficient, or the signal's side isn't scorable for its type. This is a data availability issue, not a directional miss. Undetermined signals are excluded from the win-rate denominator.

Experimental signal types are new or recently re-tuned; their thresholds are being calibrated from the very numbers shown here. Treat their win rates as provisional until sample size grows and the status flips to Validated.

Episode-aware performance (V2)

A signal often re-fires on the same market while one move plays out. The row-based table above counts each of those re-fires as an independent result; this view collapses correlated re-fires into episodes and scores each episode once, so one move is not counted many times. It always covers the full 45-day window, which spans the three disjoint validation periods (0–7, 7–21 and 21–45 days ago) behind the stability verdict.

Loading episode-aware performance…

Measurement limitations. Returns are measured market moves over each signal's scoring horizon, using that type's outcome method — not executed trades. They exclude fees, slippage, fills, position sizing and execution, and they do not describe the result of any trade. Past episodes say nothing certain about future ones.

Win rate (all scored) counts flats as non-wins; its Wilson 95% interval belongs to that rate. Decisive win rate ignores flats and has no interval here. CVaR 5% is the mean of the worst 5% of episode returns. The mean-return interval is a bootstrap over episodes and is withheld when the sample is too small.

Proposed status is what this methodology would assign. It is not the production calibration status shown in the row-based table, and it changes nothing about how signals fire.