Accuracy Methodology

The headline win rate covers signals that resolved in the last 30 days. It is a rolling window, recomputed continuously, not a hand-picked period. A signal counts toward the number on the day it closes, not the day it was issued, so the metric always reflects recently-settled outcomes rather than a frozen historical best.

Every directional signal ships with a concrete plan: an entry, a stop loss, and one or more targets. We do not grade a signal on whether the asset "eventually" moved the right way. We replay the real price path after the signal was issued and let it hit a stop or a target the same way a live position would:

• Price reaches the target before the stop → the signal closes as a win.
• Price reaches the stop first → the signal closes as a loss.
• The signal expires still open → it is settled at its expiry price, counted as a profit or a loss by where it actually landed.

The realized outcome is recorded as net PnL, after trading fees and any leverage — not the raw asset move. A small favorable move that would not cover fees is not quietly booked as a win.

Each closed signal lands in exactly one bucket, keyed on the sign of its realized net PnL so the published win rate stays consistent with the published average PnL:

The denominator is resolved signals only — wins plus losses. Flat and excluded signals are in neither the numerator nor the denominator. The audit finding that drove this page was that losses had been invisible; every aggregate we publish now reports the loss count explicitly, right next to the wins.

This is deliberately the stricter of the two metrics we could publish. A looser definition — "did the asset move our direction at all over 24 hours" — would inflate the number by counting almost any wiggle as a win. We do not use that metric for any public number.

Every signal also carries a confidence. Calibration asks the only question that makes a confidence number meaningful: when we say 80% confident, do those signals actually win about 80% of the time?

We bucket resolved signals by their stated confidence (60–64, 65–69, …, 95–100) and, for each bucket, compute the realized win rate from section 4. The edge is the gap between the two:

A positive edge means a bucket out-performed its own confidence; a negative edge means it was over-confident. We publish the full calibration table, including negative edges, on the track record. Shipping a table that shows where the model was over-confident is more credible than not shipping one — so we show it and keep tuning.

Small samples lie. A 100% win rate over three trades is noise, not a track record. Two safeguards keep the numbers honest:

• The customer-signal win rate is only surfaced as a headline once the resolved sample is large enough to be defensible. Below that we lead with the on-chain Polymarket record instead.
• Calibration buckets with fewer than 25 resolved signals are flagged low so a thin bucket is never read as settled fact. Platform-wide win rate is reported with a Wilson 95% confidence interval so you can see the statistical range, not just the point estimate.

None of this asks for trust. The signal feed is public, the calibration table is public, and the autonomous Polymarket bot trades real USDC from a wallet you can audit on-chain.

• The live 30-day win rate, calibration buckets, and worst trades: /track-record
• Every signal broken down by domain, timeframe, and token: /leaderboard