Trading Methodology

Vauban Finance combines deterministic market structure analysis (ICT framework) with multi-timeframe deep learning to generate high-conviction trading signals. Every model is continuously validated against out-of-sample data before serving predictions.

1. ML Architecture — VxLSTM Ensemble

Our prediction engine uses a Multi-Timeframe Encoder architecture: 9 independent encoders process OHLCV data across timeframes from 1-minute to daily bars, each producing a 48-feature representation. Cross-Timeframe Attention fuses these into a unified latent space before direction prediction.

Reference: xLSTM architecture based on arxiv:2405.04517 (Beck et al., 2024). Mamba SSM: arxiv:2312.00752.

2. ICT Market Structure

The Inner Circle Trader (ICT) framework provides deterministic pattern detection: Order Blocks (OB), Fair Value Gaps (FVG), Market Structure Shifts (MSS), and Liquidity Sweeps identify high-probability entry zones. ICT signals are used both as standalone indicators and as features for the ML ensemble.

3. Walk-Forward Validation

Models are only served when they pass continuous walk-forward validation. The ML Gate (TRADING_ML_GATE_OPEN) remains closed until Sharpe ratio exceeds 1.0 on 6 months of out-of-sample data. Live performance is monitored with a 7-day rolling Sharpe — automatic fallback to the previous model if degradation is detected.

4. Macro Regime Overlay

Gavekal fractal regime detection (Growth/Inflation 2×2 matrix) combined with a Hidden Markov Model provides macro context. Predictions are regime-weighted: conviction is adjusted by the current regime's historical edge for the signal type.