AI Research

Teaching a machine
to read the market.

We are building a Temporal Fusion Transformer trained on live tick, orderbook, and candle data from futures markets. This is long-term, open research. We document everything as we go.

Phase 01 active — live data collection running now

The Architecture

Temporal Fusion
Transformer

The TFT was built at Google DeepMind for exactly this kind of problem — forecasting complex, multi-variate time series where different inputs matter at different times.

Unlike a black-box neural network, the TFT uses variable selection networks and multi-head attention to show you exactly which inputs it weighted most heavily for each prediction. Every decision is interpretable.

LSTM Encoder
Processes sequences of tick and candle data to capture temporal patterns
Multi-Head Attention
Learns which time steps and features are most predictive for each forecast
Variable Selection
Automatically weights orderbook, tick, and candle inputs by importance
Quantile Outputs
Predicts a full probability distribution — not just a single price target
Multi-Horizon
Simultaneously forecasts 1min, 5min, 15min price displacement

The Roadmap

01Live Now

Data Collection

Tick-level trade data, full orderbook depth (bid/ask ladders), and multi-timeframe candle data collected continuously from live futures markets. Every microsecond of market activity is captured, timestamped, and stored. This is the raw material the model will train on.

Tick
Resolution
24/7
Collection
Live
Markets
02In Progress

Feature Engineering

Raw market data is transformed into meaningful signals — orderbook imbalance, trade flow toxicity, bid/ask pressure, volume delta, price displacement relative to VWAP. The model learns which features matter at which moments in time.

OB
Imbalance
Flow
Toxicity
VWAP
Delta
03Next

Model Training

A Temporal Fusion Transformer (TFT) — an architecture built specifically for multi-horizon time series forecasting. It combines LSTM sequence encoders with multi-head attention, learning which inputs are predictive at each point in time. No black box: every attention weight is interpretable.

TFT
Architecture
Multi
Horizon
100%
Interpretable
04Next

Signal Generation

The trained model outputs a probability distribution over future price displacement — not a single prediction, but a range of outcomes with confidence intervals. These signals are used to filter and validate existing rule-based strategies before execution.

Prob.
Distribution
CI
Intervals
Filter
Layer
05Future

Live Deployment

The validated model runs alongside the existing bot infrastructure, providing a continuous AI signal layer. Over time, as confidence in the model grows, it takes on more of the decision making — evolving from filter to co-pilot to primary signal source.

Live
Inference
Co-pilot
Mode
Evolving
Role

Follow the Research

We document everything.

Every experiment, every result, every failure — published openly as the research progresses.

Follow the Journey →