Every M2M card carries three confidence numbers, not one. A single blended number hides the most decision-relevant signal — which dimension of confidence is high versus low.
Most signal services treat confidence as something an analyst assigns based on intuition. That framing breaks the moment the prediction turns binary: a trade labeled “75% confident” that loses doesn't mean the system was wrong — it means the system claimed probability, and one sample doesn't refute a probability claim.
M2M inverts this. Confidence is a property of the discipline: when we say 75%, the only thing we claim is “over many trades like this, roughly 75% win.” Our weekly calibration audit tests whether that claim holds. A system whose 75% cards win 73–78% over 60+ outcomes is genuinely calibrated. One whose 75% cards win 55% is broken, regardless of how confident it sounds.
A mechanical calculation: given the trade structure, current price, implied volatility, and time to expiration, what is the probability the position finishes profitable at expiry? This uses standard options-pricing math and does not incorporate any edge claim — it assumes the market is fairly priced.
A credit spread with a 0.30-delta short strike has roughly 70% POP. A long call at the money might be 35–45%. These are structural properties of the trade, not opinions.
The system's confidence that it identified a genuine informational edge, separate from POP. This integrates convergence across specialists (are multiple independent analyses agreeing?), backtest sample size, regime alignment, citation quality, and penalty deductions for stale data or redundant signals.
Two trades with identical Confidence A can have very different Confidence B — one might have three specialists converging with strong backtests, while the other has a single weak signal. This is the number our calibration audits primarily test.
Given Confidence A (probability), the structure's risk/reward, and Confidence B (likelihood the edge claim is real), does this trade make money in expectation after frictions? Confidence C goes to zero when expected value is negative even if A and B are high — a structure with high POP and strong edge confidence but unfavorable max-loss relative to frictions has C = 0.
| A | B | C | What it means |
|---|---|---|---|
| High | High | High | Strong candidate — all dimensions aligned |
| High | Low | Low | Good structure, weak evidence — POP alone doesn't earn publication |
| Low | High | High | Asymmetric — low POP but real edge, positive EV from convexity |
| High | High | Low | Edge is real but frictions destroy EV — suppressed |
No confidence number on any M2M card exceeds 88%. This is a deliberate hard cap, not a bug. Options markets contain irreducible uncertainty — gap risk, liquidity withdrawal, correlation breakdown — that no retail-data analysis can eliminate. Displaying 95% would be dishonest.
Additionally, sample-size caps apply: strategies with fewer than 30 backtested trades cannot exceed 60% confidence, fewer than 100 cannot exceed 75%, and fewer than 500 cannot exceed 85%. Small samples produce noisy estimates — the caps enforce humility until the data earns higher claims.
A 75%-confident trade that loses is not evidence the system is broken. It is one sample from a distribution. A coin that comes up tails once is not evidence it's unfair. Confidence claims are testable only over many outcomes — which is exactly what M2M's calibration audit does every week.
The right question is never “did this one trade win?” but “over the last 60 trades in this confidence band, does the realized win rate match the predicted confidence?” If it does, the system is calibrated. If it doesn't, the system rolls back to shadow mode automatically until the gap is investigated.
Every edge claim on an M2M card traces to one of two sources:
Curated research entries with tier ratings. Each entry cites its source (academic paper, exchange data, practitioner study) and states what it does and does not claim.
Walk-forward validated strategy results. Parameters optimized on in-sample data, tested on unseen out-of-sample data. Strategies that degrade significantly between periods are shelved, not published.
If a card cannot cite either source for its edge claim, it routes to watchlist with no confidence assigned.