Signal Quality Classification
An ML-based system that evaluates each trade signal on a 5-tier scale (A through E) based on historical performance of similar setups, providing an additional quality filter beyond raw technical criteria.
See Signal Quality Classification in real trade signals
Tradewink uses signal quality classification as part of its AI signal pipeline. Get signals with full analysis — free to start.
Explained Simply
Not all signals that meet technical entry criteria are equal. A breakout on a quiet Tuesday with average volume is a very different quality setup than a breakout with 3x average volume on a day when the sector is rotating strongly. Signal quality classification captures these contextual differences.
The classifier is trained on historical trade outcomes. It learns which combinations of features — pattern type, volume ratio, sector momentum, time of day, market regime, ATR expansion, relative strength vs. sector — correlate with strong vs. poor subsequent performance.
The 5 tiers:
- A-grade: Highest historical win rate and R-multiple in similar conditions. Execute at full size.
- B-grade: Above-average quality. Standard position sizing.
- C-grade: Average. Acceptable setups that can be taken or skipped based on conviction score.
- D-grade: Below average. Generally skipped or reduced to half size.
- E-grade: Historically poor performance in similar setups. Filtered out regardless of technical appearance.
What Makes a Signal High Quality?
High-quality signals share several empirically validated characteristics. Volume confirmation is critical: breakouts with 2x or more average relative volume have significantly higher follow-through rates than low-volume breakouts that are easily reversed. Sector alignment matters: a stock breaking out while its sector ETF is also trending up has a fundamentally different quality profile than an isolated stock move against sector headwinds. Market regime fit is essential — the same technical pattern has different expected outcomes in trending versus choppy regimes. Time of day is a powerful quality signal: the first 30 minutes and last 30 minutes of the session have the highest directional follow-through; midday low-volume setups carry higher reversal risk. A quality classifier learns to weight these contextual factors as a unit rather than applying them as sequential filters.
ML Features Used in Signal Classification
Effective signal quality classifiers use a combination of signal-level features (the setup itself) and market-context features (the environment). Signal features include: pattern type (breakout, VWAP reclaim, ORB, etc.), ATR percentile at entry (is volatility elevated or compressed), RSI at entry, and whether the entry is above or below VWAP. Context features include: relative volume versus 20-day average, sector ETF momentum, VIX level and its recent change, time since market open (minutes into session), and intraday regime classification. The combination of these features allows the classifier to distinguish setups that look identical technically but have very different historical win rates based on their surrounding context.
Training and Retraining the Classifier
Signal quality classifiers degrade over time as market microstructure evolves and trading patterns change. A model trained on 2022 data may not reflect the dynamics of 2025 markets. Walk-forward retraining addresses this: after each trading month, new trade outcomes are added to the training set and the model is retrained on the most recent N months of data (a rolling window). This keeps the classifier current without discarding all historical context. Tradewink's MLRetrainer handles this automatically — as trade outcomes accumulate in the database, the classifier is periodically retrained and A/B tested against the current production model before being promoted. Features that lose predictive power over time are downweighted; newly informative features are added.
Integrating Signal Quality with Conviction Scoring
Signal quality classification and conviction scoring address complementary dimensions. The signal quality classifier asks: 'Given what we know about this setup type and market context, how have similar setups historically performed?' It answers with a historical-pattern grade (A through E). Conviction scoring asks: 'Given the current AI analysis of this specific ticker right now — news, sentiment, technical alignment, trade lessons — how confident is the system in this particular trade?' It answers with a 0–100 score. Both signals are used together: a high conviction score on a D-grade signal still gets reduced position sizing. A C-grade signal with very high conviction may get executed at standard size. Neither dimension alone is sufficient — combining both produces more robust execution decisions than either independently.
How to Use Signal Quality Classification
- 1
Define Quality Tiers
Classify signals into A (highest quality), B (good), and C (marginal). A signals: multiple confirming indicators, volume confirmation, regime alignment, and clean chart pattern. B signals: 2-3 confirmations. C signals: single indicator with no additional confirmation.
- 2
Map Actions to Signal Quality
A signals: take full position, aggressive entries. B signals: reduced position size, wait for pullback entry. C signals: skip or paper trade only. This mapping ensures you allocate capital proportional to signal quality.
- 3
Train a Classification Model
Build a simple ML model (logistic regression or random forest) using historical signal features (indicator values, volume, regime) to predict trade outcomes. Use the model's probability output as the signal quality score. Recalibrate quarterly on new data.
Frequently Asked Questions
What is the difference between A-grade and B-grade signals?
A-grade signals are in the top historical performance tier for setups with similar features and market context — typically the top 15–20% by expected win rate and R-multiple. B-grade signals are above-average but not top-tier. In practice, the difference matters most for position sizing: A-grade signals get full position size, B-grade signals get standard size, and C-grade signals may get reduced size or be skipped during marginal market conditions. The grade boundaries are determined by the ML model's learned thresholds, not arbitrary cutoffs.
Can a signal quality classifier be used without ML?
Yes. A rule-based classification system — for example, requiring volume above 1.5x average, sector ETF trend positive, VIX below 25, and time of day within power hours — is a rudimentary quality classifier. It is more interpretable but less adaptive than an ML model. Rule-based classifiers cannot discover non-obvious interactions between features (e.g., volume is more predictive when combined with specific ATR conditions) that ML models learn automatically from data.
How many trades does it take to build a reliable signal classifier?
Reliable training requires several hundred to a few thousand labeled trade examples per signal type. With fewer examples, the classifier overfits to noise in the training data. For a system trading 3–10 positions per day, accumulating 500–1,000 trades per strategy takes 3–6 months. During this bootstrapping period, simpler rule-based quality filters are more reliable than ML classifiers. Tradewink initializes new strategies with a conservative rule-based filter and transitions to the ML classifier once sufficient training data has accumulated.
Does Tradewink automatically retrain the signal classifier?
Yes. The MLRetrainer handles automatic retraining on a scheduled cadence as new trade outcomes accumulate. Each retrain uses a rolling window of recent data to keep the classifier current with evolving market conditions. The retrained model is A/B tested against the current production model on held-out data before being promoted — preventing a worse model from being deployed. Retraining results are logged for audit and review.
How Tradewink Uses Signal Quality Classification
The SignalClassifier in Tradewink processes each candidate through an ML model trained on thousands of historical trade outcomes. Features include: signal type, timeframe, volume ratio, sector ETF performance, VIX level, time since market open, ATR percentile, and RSI at entry. The classifier outputs a quality tier that combines with conviction score and composite rank to determine final execution decision. The classifier is retrained automatically via the MLRetrainer as new trade outcomes accumulate, keeping the model current with changing market conditions.
Trading Insights Newsletter
Weekly deep-dives on strategy, signals, and market structure — written for active traders. No spam, unsubscribe anytime.
Related Terms
Learn More
AI Day Trading Strategies: The Complete 2026 Guide
A complete guide to AI-powered day trading strategies in 2026. Learn how artificial intelligence applies breakout, mean reversion, VWAP, and options strategies — and how autonomous agents like Tradewink execute the full pipeline from screening to exit.
How AI Day Trading Bots Actually Work: The 8-Stage Pipeline from Data to Execution
A builder's breakdown of a production AI day trading system. Covers the full pipeline: market data ingestion, regime detection, screening, AI conviction scoring, position sizing, execution, dynamic exits, and self-improvement.
What Is AI Trading? A Complete Guide for 2026
AI trading uses artificial intelligence to analyze markets, identify opportunities, and execute trades. Learn how it works, its advantages over manual trading, and how to get started.
How AI Trading Signals Work: From Data to Trade Idea
Ever wonder how AI generates trading signals? We break down the full pipeline: data ingestion, pattern recognition, scoring, filtering, and delivery.
See Signal Quality Classification in real trade signals
Tradewink uses signal quality classification as part of its AI signal pipeline. Get daily trade ideas with full analysis — free to start.