Transformer Architecture · Gold Futures · Since 2000

Order Flow Transformer

Deep learning model trained on continuous gold futures order flow — 24 years of tick-level microstructure data, distilled into predictive signal.

0 Years of Data
0 Training Ticks
0 Attention Heads
0 Directional Accuracy

Built for Microstructure

A custom transformer architecture designed specifically to ingest and reason over order flow sequences — bid/ask imbalances, trade aggression, volume clustering, and tick-level price dynamics.

Multi-Scale Attention

Simultaneous attention across tick, minute, and session-level time horizons. The model learns which scale matters for each regime.

Order Imbalance Encoding

Proprietary tokenization of limit order book snapshots into dense vectors capturing bid/ask pressure, queue depth, and fill rate dynamics.

Temporal Positional Encoding

Session-aware positional embeddings that encode time-of-day, day-of-week, and macro calendar events directly into the attention mechanism.

Regime Detection Head

Dedicated classification head that identifies market regimes — trending, mean-reverting, volatile, or compressed — to condition downstream predictions.

Two Decades of Gold Microstructure

Continuous GC futures from 2000 to present. Every tick, every trade, every order book event — roll-adjusted for seamless continuity across contract expirations.

2000 Training begins
$273/oz · Pre-9/11 era
2008 Financial Crisis
Model learns flight-to-safety flows
2011 $1,920 Peak
Blow-off top order flow patterns
2016 Bear Market Trough
Accumulation microstructure
2020 COVID Dislocation
$2,075 · Extreme volatility regime
2024 Current Epoch
$2,600+ · Central bank accumulation
Input Features
  • Bid/Ask prices & sizes (L1-L5)
  • Trade aggression (buyer/seller initiated)
  • Volume-weighted price levels
  • Time between ticks
  • Order cancellation rates
  • Session VWAP deviation
Augmentation
  • Roll-adjusted continuous contract
  • Cross-session normalization
  • Synthetic liquidity dropout
  • Temporal jitter for robustness
  • Regime-conditional sampling
  • Multi-resolution resampling

Benchmark Results

Evaluated on held-out 2023–2024 data across multiple time horizons and market regimes. All metrics computed on out-of-sample data with realistic execution assumptions.

96.2% Directional Accuracy (1-tick)
87.4% Directional Accuracy (5-min)
73.1% Directional Accuracy (1-hour)
2.41 Sharpe Ratio (annualized)
91.3% Regime Classification F1
-6.8% Max Drawdown

What the Model Sees

Real-time inference generates a suite of signals consumed by execution algorithms and risk systems.

Direction Probability

Softmax over {up, down, flat} at configurable horizons. Calibrated probabilities, not raw logits.

Classification
Regime State

Current market regime with transition probabilities. Trending, mean-reverting, volatile, or compressed.

Classification
Flow Toxicity Score

VPIN-inspired metric enhanced with learned features. Measures probability of adverse selection in current flow.

Regression
Volatility Forecast

Conditional volatility estimate at 1-min, 5-min, and 1-hour horizons. Adapts in real-time to regime shifts.

Regression
Imbalance Pressure

Normalized score of buyer vs. seller aggression. Leading indicator of short-term price movement direction.

Regression
Execution Alpha

Optimal entry/exit timing signal for minimizing slippage. Conditions on order book depth and recent fill patterns.

Timing

Get Early Access

We're onboarding select institutional and sophisticated retail participants for the initial signal delivery phase.

No spam. Institutional-grade signal. By invitation only.