Skip to content

Tessera

Mid-frequency ML trading system for crypto perpetual futures.

Tessera is a research-grade, fully reproducible trading system for USDT-margined perpetuals on Binance and Bybit. It covers every layer of the ML trading stack — from raw market data to live paper trading — with rigorous statistical validation at each step.

Headline Results

Metric Value
Backtest Sharpe (walk-forward) 1.41
Deflated Sharpe (N=247 trials) 0.87
95 % bootstrap CI [0.52, 1.19]
Paper-trading Sharpe (48 h) 1.31
Max drawdown (4 years) 8.2 %

Quick Start

make setup      # uv sync --all-extras + pre-commit install
make lint       # ruff + mypy --strict
make test       # pytest with coverage
make backtest   # full walk-forward evaluation
make figures    # regenerate all paper and docs figures
make docs       # serve this site at http://localhost:8000

Pipeline Overview

Exchange WebSocket
        │
        ▼
  CCXT Ingestor ──── 1-min OHLCV + funding rates
        │                          │
        ▼                          │
  Parquet Store                    │
  (PyArrow, partitioned)           │
        │                          │
        ▼                          ▼
  Feature Pipeline ◄─── 20 features across 6 families
  (topo-sorted, PIT-safe,          (microstructure, vol,
   per-day Parquet cache)           funding, cross-sectional,
        │                           returns, regime)
        ▼
  Triple-Barrier Labeler
  (σ-scaled, uniqueness-weighted samples)
        │
        ▼
  Cross-Validation
  (PurgedKFold → CPCV 15-split → WalkForward)
        │
        ├── LightGBM Primary (200 Optuna trials)
        │
        ├── LightGBM Meta (OOS predictions only)
        │
        └── Ensemble → Quarter-Kelly sizing → Vol-target
                │
                ▼
          Risk Stack
          (circuit breakers · kill switches · limits)
                │
                ▼
          Nautilus Trader
          (fees · slippage · latency model)
                │
                ▼
          Exchange API  ←→  Prometheus → Grafana
Page What you'll find
Architecture Detailed component diagram, config and logging layers, Docker Compose setup
Methodology Triple-barrier labeling, CPCV, deflated Sharpe, bootstrap CI, stress windows
Features All 20 features with formulas and references
Models Model cards: LightGBM, PatchTST, Chronos, ensemble
Results Tear sheets, ablation tables, stress-window analysis
Runbook Kill-switch and circuit-breaker incident response
Pitfalls Every look-ahead leak and data bug found and fixed

Research Paper

The full methodology is documented in a 12-page IEEE-style paper: paper/main.tex

Compile with:

cd paper
pdflatex main.tex && bibtex main && pdflatex main.tex && pdflatex main.tex

Components

Module Description
tessera.data CCXT ingestor, Parquet store, universe, validation
tessera.features 20 features: microstructure, vol, funding, cross-sectional, regime
tessera.labels Triple-barrier labeler (AFML §3), sample weights (AFML §4)
tessera.cv PurgedKFold, CPCV, WalkForward CV
tessera.models LightGBM, PatchTST, Chronos, TFT, ensemble, model registry
tessera.backtest Nautilus engine, fee model, square-root slippage
tessera.risk Quarter-Kelly, vol-target, limits, circuit breaker, kill switch
tessera.live PaperRunner, healthcheck, position reconciliation
tessera.strategies MLDirectionalStrategy

Citation

@software{lunawat2026tessera,
  author  = {Lunawat, Yash},
  title   = {Tessera: A Mid-Frequency ML Trading System for Crypto Perpetual Futures},
  year    = {2026},
  url     = {https://github.com/Yash121l/Tessera}
}