Conviction Scoring

• 80–100 — Very high conviction: strong multi-factor alignment, trade the full position size
• 60–79 — Standard conviction: trade at normal position size, signal meets all criteria
• 40–59 — Marginal conviction: reduced position size or skip depending on market conditions
• Below 40 — Low conviction: signal is filtered out, no trade executed

Higher conviction does not guarantee a winning trade — markets are probabilistic. But higher-conviction trades should, over a large sample, have better win rates and risk/reward than lower-conviction ones. This relationship is monitored by the ConfidenceCalibrator, which detects when the conviction model is overconfident (predicting 80+ but actually winning at 50% rates) and recalibrates the scoring weights.

For high-stakes trade decisions, Tradewink's 3-agent team evaluation provides a more rigorous conviction assessment than any single AI call:

Bull analyst: Reviews the trade case through an optimistic lens. Focuses on technical breakout quality, momentum signals, sector alignment, positive sentiment catalysts, and why the setup meets entry criteria. The bull analyst's conviction is anchored to the strongest supporting evidence.

Bear analyst: Reviews the same setup adversarially. Looks for reasons the trade should fail — unfavorable market regime, deteriorating strategy health, weak volume confirmation, opposing macro headwinds, or prior instances where similar setups failed. The bear analyst's job is to find holes in the thesis.

Meta-analyst (debate moderator): Reads both reports, weighs the strength of each argument, and synthesizes a final conviction score. The meta-analyst is not biased toward either direction — it asks: given the bull case and the bear case, how strong is the overall evidence for this trade? The final conviction score reflects the balance of evidence after both sides have been heard.

This structure is more reliable than a single-agent evaluation because it forces the system to surface counterarguments rather than rationalizing a preexisting bullish or bearish inclination. Research on structured debate in expert forecasting (superforecasting) consistently shows that adversarial deliberation reduces overconfidence and improves calibration. Tradewink applies this principle to trade evaluation.

A conviction score is only useful if it is calibrated — meaning that trades scored at 80 conviction should actually win more often than trades scored at 60, and the win rate gap should be statistically meaningful.

Overconfidence: The most common failure mode. An uncalibrated model consistently scores trades at 75-90 when the actual win rate at those scores is only 50-55%. Overconfidence leads to overtrading — the system executes too many marginal setups because they appear high conviction.

Underconfidence: Rarer but equally problematic. The model scores trades at 40-55 when actual win rates are 65%+. Underconfidence leaves profitable setups on the table.

Calibration curve: A properly calibrated system shows a monotonic relationship between conviction score and realized win rate. Score 40-50 should have ~45% win rate. Score 60-70 should have ~60% win rate. Score 80+ should have ~75%+ win rate. If the actual curve is flat or inverted in any range, the model is miscalibrated for that range.

The ConfidenceCalibrator: Tradewink's ConfidenceCalibrator runs periodically and compares conviction score distributions against realized trade outcomes. When the calibration curve deviates significantly from the ideal shape, it adjusts the scoring weights for the components that are most over or under-weighted. This feedback loop ensures the conviction score remains a meaningful predictor of trade quality rather than a number that degrades over time.