StatBit Alpha builds high- and mid-frequency trading strategies with state-of-the-art deep learning, statistical machine learning and mathematical models, and tests every one of them as if trying to break it.
We work at two ends of the time spectrum. The data, models and engineering change completely between them, but the research discipline behind both is the same.
Strategies that hold positions for minutes to days, built on signals that are statistically robust rather than merely fast.
A low-latency stack that reads the order book directly and reacts in milliseconds. It is rolling out soon.
Drag the slider to see how the data, models and hardest problem change as the holding period gets longer.
We don't pick a model because it's fashionable. Each one earns its place by beating simpler alternatives on data it has never seen.
Attention-based architectures for multi-horizon forecasting, learning which parts of market history matter for what comes next.
Convolutional and recurrent networks that learn features directly from raw limit order book snapshots instead of hand-crafted indicators.
Fast, parallel sequence models suited to the tight inference budgets of short-horizon trading.
Pre-training on large volumes of unlabelled market data so downstream models start from a richer view of market state.
Strong, interpretable baselines for tabular market features and cross-sectional ranking.
Simple models with strict regularisation. Often hard to beat, and the bar every complex model must clear.
Combining diverse signals so no single model's failure becomes the portfolio's failure.
Models that update as new data arrives, because markets change their behaviour over time.
Mean-reverting and jump-diffusion models for prices, spreads and volatility.
Finding instruments whose prices are tied together over the long run, and trading the temporary gaps.
Kalman filters and hidden Markov models to track hidden market regimes and changing relationships.
Market-impact-aware execution and portfolio optimisation under explicit risk constraints.
Every idea follows the same path. Most never make it past validation, and that is the point.
Clean, versioned market datasets with full lineage, so every result can be reproduced.
Signals engineered from prices, volumes and order flow, plus features learned by deep models.
Candidates from all three model families compete against simple baselines.
Walk-forward tests, purged cross-validation, realistic costs and crisis stress tests.
The strategy runs live on real-time data in a production-identical environment before any capital is committed.
Live behaviour is compared against test expectations, with automatic alerts on any drift.
A backtest that looks perfect is usually a warning sign. Our testing is built to find the reasons a strategy will fail before the market does.
Try it: why validation matters
Increase the model's complexity. It fits the past better and better, then reveal data it has never seen.
Every strategy we deploy runs inside the same risk framework. It decides how much we can lose before any model decides how much we might make.
Stress test our risk controls
Pick a market shock and watch the controls respond.
Simulated illustration of how the controls behave. Not actual trading results.
A cloud-native stack lets us move from an idea to a fully tested strategy quickly, and scale compute when a problem demands it.
Elastic CPU and GPU capacity on AWS for large backtests and deep learning training runs.
PyTorch for deep learning, Polars for fast columnar data processing, and scikit-learn for classical models.
Versioned data, tracked experiments and code under version control, so any result can be rebuilt.
A purpose-built execution layer for our high-frequency stack, currently in development.
We're early, technical and building every layer ourselves: research, infrastructure and risk. That's the best time to get involved. Follow our progress on LinkedIn.
If you back quantitative teams at the earliest stage, we'd like to walk you through our research process, our risk framework and our roadmap for mid- and high-frequency trading.
If you think in distributions, enjoy hard data problems and want your models to meet real markets, we want to hear from you.