Strategy AI picks which strategy to deploy. Signal Score AI sits inside the strategies you've already deployed and judges each individual signal — on eight dimensions, against a strategy-specific threshold tied to the risk profile you set. Bad signals are refused before they become orders. The scoring weights adapt to your real outcomes over time, so the system you run after a hundred trades is genuinely tuned to your account.
Each step happens in milliseconds inside the strategy loop. The scoring is automatic — no per-trade dashboard click required.
An emitting strategy (RSI, MACD, MA Cross, Momentum, Mean Reversion, Sentiment, TradingView) flags a potential entry on candle close.
The scorer evaluates the candidate across all enabled factors. Each factor is 0–100; the final score is a weighted sum, normalised to 0–100.
The final score is compared to the threshold for this strategy at your configured risk level. Above the gate, the signal proceeds. Below, it's logged and skipped.
When the resulting trade exits, the realised P&L is fed back to the scorer. Over time the weights adjust to favour factors that have actually worked on your account.
The default profile weights the most evidence-backed factors highest — support/resistance, trend, and volume. Sentiment and candlestick patterns are off by default; you can opt them in.
The factor weights and thresholds differ per strategy because the strategies look for different things. The defaults reflect that on day one — they get tighter over time as the scorer learns from your outcomes.
RSI, MACD, MA Cross, Momentum, Mean Reversion, TradingView — each has its own factor weights and confidence threshold. A signal that's high-conviction for one strategy doesn't necessarily pass for another.
A strategy's own indicator is dropped from the scorer's indicator factor — so RSI isn't asked to confirm RSI, and MACD isn't asked to confirm MACD. Independent confirmation only.
LOW, MEDIUM, HIGH, and VERY_HIGH risk profiles each have a different threshold per strategy. Higher risk lowers the gate so more signals pass — fewer trades skipped, but lower average conviction per trade.
Mean Reversion is contrarian — it already filters with its own ADX and RSI checks. Asking the scorer to also penalise trend disagreement would double-judge the signal. The trend and indicator factors are deliberately removed from MR's scoring math; the remaining factors are reweighted to fill the gap.
Signal Score AI only runs on strategies that emit a discrete trade signal. Strategies that work by placing orders at intervals rather than reacting to a fresh signal do not flow through the scorer — they are not gated by it and not learned from.
When a trade closes, the scorer records what each factor scored at entry against whether the trade made or lost money. After enough trades, factors that have actually predicted profit on your account get more weight; factors that have not get less.
Only completed exits feed learning — and only the ones with normal exit reasons (signal, SL, TP). Manual closes, emergency stops, liquidations, and timeouts are filtered out so the scorer learns from your strategy, not your interventions.
Global weights apply from day one. After ~50 trades on a given strategy type, type-level weights start blending in. After ~100 trades on a specific strategy instance, instance-level weights blend in further. New accounts start with the production-default weights, not zero.
Every factor weight is bounded between sensible floor and ceiling values so no single factor can ever dominate the score, and the total is renormalised to one after each update.
For each signal the strategy considered, the dashboard shows each factor's score, the weighted final, the threshold it was compared against, and whether the signal was ACCEPTED or REJECTED.
Strategy-level acceptance rates and win rates over the periods you choose, so you can see which strategies the scorer is helping versus which it isn't.
Each scored signal is annotated with the market regime it fired in — trending vs. ranging, volatility level — so you can see whether bad scores cluster in specific regimes.
Every refused signal is logged with the specific reason it failed — which factor scored low, which threshold it missed. Useful when a strategy seems quiet and you want to know whether it's seeing nothing or seeing things and refusing them.
A few concrete boundaries worth knowing so you can read acceptance and rejection behaviour correctly.
The scorer learns from every completed trade rather than holding out a recent window as a test set. That keeps the sample size as large as possible; the tradeoff is that recent drift only shows up in the weights once enough new trades have accumulated to readjust.
Learning is per-strategy and per-strategy-instance. A BTC/USDT and an ETH/USDT instance of the same RSI strategy share the same learned weights. Symbol-level granularity is on the research backlog.
The final score is a weighted sum of the eight factors. That's transparent and auditable — every rejection points to a specific factor. Interaction-aware or gradient-boosted models (where pairs of factors reinforce each other) are on the research backlog rather than in production.
Multi-timeframe consensus checks the 1-hour, 4-hour, and 24-hour timeframes. Strategies running on 5-minute or 15-minute bars still get the scorer — they just don't get the sub-hourly consensus factor's lift.
Factor weights adapt from real trade P&L rather than user-set overrides — the model is what your account has actually earned, not what you thought would work. An admin endpoint exists for ops if a regime change demands intervention.
The Signal Insights dashboard shows current weights vs. cold-start defaults, so drift is visible when you look. There is no push alert today — you notice by checking, not by getting pinged.
Signal Score AI is automatic — it runs as soon as you deploy an emitting strategy. No configuration step required.