AlgorithmSignal ScoringOpen Source

How TradeClaw Scores Trading Signals: A Full Walkthrough

March 22, 2026 · 12 min read

Most AI trading signal tools are black boxes — you get a "BUY" label and a confidence percentage, with no explanation of how it was computed. TradeClaw is different: the entire scoring engine is open source, and this post walks through exactly how every signal gets generated.

The Big Picture: A Points System

TradeClaw uses a simple, deterministic points system. Every signal is scored by running 5 technical indicators and awarding points based on what each indicator shows. The total is normalised to a confidence percentage.

IndicatorMax PointsWhat It Measures
RSI (14)20Overbought / oversold momentum
MACD (12/26/9)20Trend direction & momentum change
EMA (20/50/200)20Price vs moving average alignment
Bollinger Bands (20, 2σ)15Price at extremes of volatility range
Stochastic (14/3/3)15K/D position and crossover
Total90→ Normalised to 50–98% confidence

Step 1: Compute Both Directions

For every asset and timeframe, TradeClaw computes a BUY score anda SELL score independently. Whichever score is higher and exceeds a minimum threshold (55% confidence) becomes the signal. If neither exceeds the threshold, no signal is emitted — "no opinion" is a valid output.

Step 2: Quality Gates (Filters)

Before a signal is emitted, it must pass 4 quality checks:

  • ATR check: Average True Range must be > 0.3% of price. Signals in flat/dead markets are filtered out.
  • Bollinger bandwidth: Must be > 1%. Consolidating markets produce false signals.
  • EMA slope: At least one EMA must have non-zero slope. Zero-slope = no trend = noise.
  • Stop distance: SL must be > 0.5% from entry. Tight stops get triggered by normal noise and produce misleading results.

Step 3: Confidence Normalisation

Raw score (0–90) → confidence (50–98%). The formula:

confidence = 50 + (rawScore / 90) × 48

Why cap at 98%? No deterministic indicator combination can justify 100% certainty. Claiming 100% confidence would be dishonest.

Is It Actually Calibrated?

Good calibration means: signals at 80% confidence should win ~80% of the time. We track this on the calibration page using Expected Calibration Error (ECE). The honest answer: with synthetic seed data, calibration is not yet validated. As live signals accumulate, that page will reflect real performance.

Multi-Timeframe Confluence

TradeClaw also generates signals across H1, H4, and D1 timeframes and checks for confluence. When all three timeframes agree, confidence gets a +15% boost. When they disagree, confidence is reduced and the signal is flagged as "conflicted."

Limitations We're Honest About

  • All indicators are lagging — they react to price, they don't predict it
  • Crypto and forex have different characteristics — RSI(14) behaves differently on 1-hour BTC vs daily EUR/USD
  • The scoring weights are manually tuned, not machine-learned from backtested data (yet)
  • No fundamental analysis, news, or macro factors are considered

The algorithm is 100% open source.If you think the weights are wrong, the quality gates are too loose, or the indicators should be different — submit a PR. That's the whole point.

View source on GitHub →