Financial Disclaimer
This article is for educational and informational purposes only. Nothing here constitutes financial advice, investment recommendations, or a solicitation to trade. Trading financial instruments involves substantial risk of loss. Past performance is not indicative of future results. Machine learning models trained on historical data do not guarantee future performance.
The WA-RSI system described in the previous article produces valid signals — mathematically sound, institutionally-calibrated readings of currency strength across timeframes. But validity is not conviction. A signal can be technically correct and still lose money, because the signal alone does not account for the context in which it appears.
The Conviction Gap is the space between a signal that passes every technical filter and a signal that a professional trader would actually act on. Bridging that gap requires something rules alone cannot provide: the ability to recognise, from a large body of historical evidence, which types of signals have consistently produced outcomes and which have not.
The Insight
A signal is a reading. Conviction is the probability that the reading reflects something real about where price is going — not just where it has been.
The Dataset
122,640
Data Points
33
Months of History
28
Currency Pairs
The model was trained on 33 months of H4 bar data across 28 major and minor currency pairs — all pairs derivable from the eight major currencies. Each data point captures the WA-RSI readings for both currencies in a pair, the spread between them, the degree of timeframe alignment, session timing, and a range of volatility context features.
Why XGBoost
XGBoost (Extreme Gradient Boosting) was chosen because of how it handles the specific characteristics of financial time-series signal data. Unlike neural networks, which require large datasets and are difficult to interpret, XGBoost builds an ensemble of decision trees that can be inspected — meaning the model can be interrogated to understand which features it finds most predictive.
- Feature importance transparency: You can see exactly which inputs drive conviction scores — critical when deploying in a live trading environment.
- Resistance to overfitting: Gradient boosting with regularisation produces models that generalise better to unseen data than deep learning architectures on datasets of this size.
- Speed: XGBoost inference is fast enough to run within MetaTrader's execution environment without latency concerns.
- Robustness to feature scaling: Financial features come in very different ranges; XGBoost handles this natively.
Feature Architecture
Input Group 01
WA-RSI Readings
The weighted average RSI for each currency in the pair, computed across H1, H4, D1, and W1 — eight values per data point.
Input Group 02
Divergence Score
The spread between the two currencies' WA-RSI values. Wider divergence does not always equal higher conviction — the model learns when spread is meaningful.
Input Group 03
Timeframe Alignment
A scalar representing how consistently the signal appears across the four timeframes. Full alignment produces the highest weight in this feature group.
Input Group 04
Volatility Context
ATR-derived volatility relative to the pair's historical range. High-conviction signals in low-volatility environments behave very differently from those in volatile regimes.
Input Group 05
Session Timing
Which of the four major FX sessions is active (Sydney, Tokyo, London, New York) and the degree of session overlap — a known driver of liquidity and signal reliability.
Input Group 06
Historical Context
Whether similar WA-RSI configurations have historically resolved in the signal direction — a form of pattern memory built into the feature set.
Validation Methodology
The model was validated using a strict walk-forward approach: trained on months 1–24 of the dataset, tested on months 25–33. No test data was seen during training. This prevents the most common failure mode of ML trading systems — in-sample overfitting that produces strong backtest statistics that collapse in live conditions.
| Metric |
WA-RSI Signals (Unfiltered) |
Conviction-Filtered Signals |
| Win Rate | 56% | 67% |
| Avg Win / Avg Loss Ratio | 1.31 | 1.58 |
| Max Consecutive Losses | 9 | 5 |
| Profit Factor | 1.39 | 2.04 |
| Signal Volume (monthly avg) | ~840 | ~210 |
The conviction filter reduces signal volume by approximately 75% — and that is the point. Not every technical signal deserves capital. The model's job is to identify the subset of signals where the historical evidence for follow-through is strongest.
What the Model Learned
The most predictive features, in order of XGBoost importance scores, are: timeframe alignment (W1 and D1 agreement matters most), the spread between WA-RSI values, session timing (London open and London/New York overlap produce the most reliable signals), and volatility regime. The H1 RSI is the weakest predictor — consistent with the intuition that short-term momentum is the least reliable component of a directional view.
Conclusion
The Conviction Gap is real, and it is quantifiable. Not all signals are equal. A machine learning layer trained on 33 months of institutional-grade currency strength data does not replace analysis — it distils it, surfacing only the signals where the historical record supports acting with confidence. That is what separates a technically valid signal from one worth trading.