AI Conviction Scoring: How AI Grades Every Trade Before You Take It

The Problem with Binary Signals

Traditional trading signals are binary: buy or don't buy. This misses critical nuance. A buy signal on a stock with perfect technical setup, strong volume, bullish options flow, and favorable market regime is very different from a buy signal on a stock that barely triggered one indicator.

Conviction scoring solves this by assigning a continuous score — typically 0 to 100 — that reflects the overall quality and confidence of a trade setup. Higher conviction = larger position size, more willingness to hold through volatility. Lower conviction = smaller size or skip entirely.

The AI trading boom: The AI trading platform market is growing at 11.4% CAGR from 2026 through 2033, with the broader AI software market reaching $174 billion in 2025. AI conviction scoring is at the heart of this growth — it is the mechanism that translates raw AI analysis into actionable, risk-adjusted trade decisions at a scale no human can match.

The Components of Conviction

A robust conviction score synthesizes multiple evidence sources:

1. Technical Score (30–40% weight)

The baseline technical quality of the setup: How clean is the breakout? Is volume confirming? How many indicators align?

Inputs typically include:

• RSI positioning (not overbought at entry)
• MACD momentum direction
• VWAP relationship (above = bullish)
• Volume vs. average (relative volume > 1.5x preferred)
• Support/resistance clarity (defined levels vs. ambiguous)
• Pattern quality (clean flag vs. messy consolidation)

Each factor contributes 0–10 points to a technical sub-score. The sub-score normalizes to 0–100 and gets weighted into the composite.

2. Market Regime Score (15–20% weight)

Does the current market environment favor this type of trade?

• Trending regime + breakout setup: high conviction boost
• Choppy regime + breakout setup: major conviction penalty
• High volatility + any setup: size reduction flag

Regime alignment is multiplicative — a technically perfect setup in a choppy regime gets penalized significantly. A mediocre setup in a strongly trending regime gets a boost.

3. Options Flow Score (10–15% weight)

What are the options players doing? "Smart money" options activity often precedes price moves.

• Large unusual call sweeps above ask: bullish signal
• Heavy put buying at key strike: bearish signal
• High IV rank with directional flow: strong conviction
• Options activity confirms price direction: score boost

Not all tickers have meaningful options flow. For low-options-volume stocks, this component gets zero-weighted.

4. Sentiment Score (10–15% weight)

News sentiment, social momentum, and analyst activity:

• Breaking positive news + positive price action: high conviction
• Analyst upgrade: moderate boost
• Negative sector news contradicting the setup: conviction penalty
• Dark pool prints (large non-public block trades): can be highly significant

5. AI-Specific Factors (10–20% weight)

This is where machine learning adds unique value:

Signal Quality Classification: An ML classifier trained on thousands of historical setups assigns a quality label (Very High, High, Medium, Low) based on how similar past setups performed. A "Very High" quality label boosts conviction; "Low" is a disqualifying factor.

Historical Win Rate at This Setup Type: If momentum breakout on stocks with this RSI + volume profile has a 68% historical win rate in current regime, that base rate factors into conviction.

Trade Lessons: Post-trade AI reflection generates natural language lessons. Similar patterns are retrieved via embedding search and factored into confidence.

Multi-Agent Debate: High-Stakes Conviction Verification

For the highest-stakes decisions (large positions, volatile markets, borderline conviction scores), a single AI model's assessment isn't enough. Tradewink uses multi-agent debate:

Bull Agent: Analyzes the setup from an optimistic perspective. Identifies supporting factors: technical setup quality, tailwinds, options flow confirmation, favorable regime.

Bear Agent: Argues the opposing case. Identifies risks: overhead resistance, high short interest, recent false breakouts in this stock, regime uncertainty.

Moderator Agent: Receives both analyses, synthesizes, and produces a final conviction score with explicit reasoning. It can override the initial score in either direction based on compelling arguments.

This three-agent process takes longer and costs more compute, but dramatically reduces false-positive signals. In testing, 3-agent evaluation improves signal quality versus single-agent evaluation — particularly in filtering setups that look technically clean but have hidden structural weaknesses.

When Does 3-Agent Mode Activate?

The multi-agent evaluation is triggered when:

• Initial composite score is in the "borderline" range (40–65): unclear whether to take the trade
• Position size would exceed 5% of portfolio
• Stock has had poor recent performance (last 3 trades lost)
• Regime is uncertain (probability distribution is flat, not peaked)
• User account tier qualifies for premium AI analysis

For fast-moving opportunities (breakout happening right now), the single-agent path is used to stay within execution time budgets.

Conviction Thresholds and Position Sizing

The conviction score directly adjusts position size:

A position that would normally be $2,000 becomes $1,000 at moderate conviction — you still participate in the setup but with appropriate skepticism.

The minimum threshold to enter a trade at all (in Tradewink's live system) is 35. Below that, the AI's uncertainty is too high to justify capital allocation.

Signal Quality Classification

Separate from the conviction score, Tradewink runs a signal quality classifier — an ML model that assigns one of five labels:

• Strong Buy (>75 confidence): Rare. Significant position warranted.
• Buy (60–75): Standard positive signal.
• Hold/Neutral (40–60): No clear edge. Wait for better setup.
• Sell (25–40): Avoid long positions, consider short.
• Strong Sell (<25): Avoid or actively short if setup confirms.

The quality classifier is trained on historical signal outcomes using walk-forward validation. It uses discretized features to avoid overfitting: RSI bins (not raw values), volume relative to 20-day average (bins), pattern type categories, and regime label.

How to Build a Simple Conviction System

Even without AI, you can build a manual conviction scoring system:

1. List 5–8 factors that matter for your setup type
2. Assign a maximum point value to each (total to 100)
3. Score each trade objectively on each factor
4. Add a minimum threshold: only trade if score exceeds 50
5. Size proportionally: $1,000 × (score / 100)

Example for a momentum breakout:

• Volume > 2x average: 20 points
• Clean chart pattern: 15 points
• Strong sector momentum: 15 points
• RSI not overbought (<70): 10 points
• Market regime trending: 20 points
• News catalyst: 10 points
• Options flow confirmation: 10 points

A trade scoring 80/100 gets full size. A trade scoring 45/100 might still be interesting but with reduced size and tighter stops.

Common Conviction Mistakes

Inflating scores on high-conviction "feeling": Conviction scoring only works if it's objective. If you're pre-deciding you want to take the trade and working backward to justify it, the system fails.

Ignoring regime in the score: Market regime is often the most important factor. A technically excellent setup in a choppy regime has statistically worse outcomes than a mediocre setup in a strong trending regime.

Using the same threshold for all setup types: Momentum breakouts and mean-reversion setups have different base rates. Consider separate scoring systems or threshold calibration per setup type.

Frequently Asked Questions

How is conviction scoring different from a simple signal?

A binary signal says "buy." Conviction scoring says "buy at 72% confidence with these specific supporting and undermining factors." The score drives position sizing and gives you explicit reasoning to review after the trade.

Does higher conviction always mean better outcomes?

Over large samples, yes. Trades with conviction >70 should significantly outperform trades with conviction <40 if your scoring system is well-calibrated. Individual trades can still be random — even a 90-conviction trade fails sometimes.

How do I know if my conviction system is well-calibrated?

Track conviction score vs. actual outcome over 100+ trades. Group by score bucket (0-40, 40-60, 60-80, 80-100) and measure win rate in each bucket. A well-calibrated system shows win rate monotonically increasing with conviction. If the 80+ bucket has the same win rate as the 40-60 bucket, your high-conviction factors aren't actually predictive.