Professional, risk-sensitive AI models specialized for capital management and active trading.
Finance is a specialized discipline — one that demands exceptional scrupulousness of approach, mathematical rigor, and multi-stage objective verification.
Four commitments that govern every model we ship — and every claim we make about it.
Statistics decides, not narrative. Every change must clear pre-stated quantitative criteria; anything not proven on data is treated as nonexistent.
Training and validation are strictly separated. A result counts only on independent periods the model has never seen — and only above the passive alternative.
We optimize not the average outcome but its distribution — with an explicit price on the left tail.
Live execution is numerically identical to the research loop: exactly what was tested is what trades — verified continuously.
A public, audited record of every trade made by autonomous AI agents on liquid markets. Independent. Adversarial. Verifiable end‑to‑end.
No human is in the loop. The agent decides position, size and timing — and accepts every consequence on its own ledger.
A trained policy — reinforcement-learning, supervised, or a hybrid — that maps live market state to a continuous action distribution.
Real orders, on real venues, with real capital. PROGMATIC reconciles every fill against venue data and publishes the full record — execution is part of how each agent is verified.
Most performance pages are a marketing surface. This one is an audit trail: 1 agents, every action timestamped, every metric recomputed from raw fills. Independence, adversarial reporting and quant-grade rigour are not features here — they are the contract.
Each fill is matched against venue data. Nothing on this site is self-reported — the ledger is the source of truth.
In-sample, out-of-sample and live are reported separately and coloured on every curve. You see exactly where fitting ends and reality begins.
Sharpe, drawdown, turnover and stationarity — across every period, not a single flattering headline number.
PROGMATIC takes no capital and manages no money. We measure agents; the venues execute. The roles never mix.
Everything on this site is research and analytics. Nothing here is advice, an offer, or a solicitation.
All information is analytical material provided "as is".
No content constitutes a trading, investment or other recommendation.
Past performance does not guarantee future results.
Trading carries substantial risk of capital loss.
You make all decisions independently and bear full responsibility.
PROGMATIC is not a broker, fund manager or financial advisor.
Nine reasons machine learning is the right tool for trading and capital management.
Neural networks process thousands of parameters in parallel — prices, volumes, macro factors, cross-instrument correlations — beyond human cognitive capacity.
Each decision is the result of generalising thousands of historical situations. Not a guess — a probabilistic model trained and validated on years of data.
Statistically grounded decisions only. No fear, no greed. Removes the most expensive failure mode of discretionary trading.
Multidimensional, nonlinear dependencies imperceptible to humans. What looks like noise to a person is a signal to the network — backed by statistical correlation.
Built on the principle of least action — the same physical law that guides photons. Machine learning minimises a loss function: not guessing, optimising the path.
The most powerful, scalable and effective tool currently known to science for financial decision-making. Synthesis of calculus, probability, statistics, physics.
Solutions are optimal under given conditions: maximise return, minimise drawdown, balance objectives. Trade-offs are explicit and tunable.
Drawdown caps, return ratios and risk constraints live inside the objective function. AI agents operate within strictly defined risk profiles by design.
Law of large numbers applied to hundreds of instruments simultaneously: deviation from expected outcomes shrinks as portfolio size grows.
6 chapters of working reference material. Each chapter on the left, its articles on the right.
Why AI is the right tool for capital management.
Agents that learn directly from rewards.
Agents that learn directly from rewards: the framework, policies and value functions.
Proximal Policy Optimisation — stable, sample-efficient policy gradients.
Generalised Advantage Estimation — variance reduction in policy gradient methods.
Building experience: how rollouts are sampled, stored and replayed during training.
Tuning λ in GAE: the bias-variance trade-off in advantage estimates.
When to scale from one agent to a multi-agent system, and what changes in training.
Multi-horizon quantile-regression TD3: a critic with two orthogonal axes — the return-distribution quantiles and the n-step bootstrap horizons.
Architecture, attention, embeddings.
Self-attention as the foundation of modern sequence models.
Layers, residuals, positional encoding — Transformers in detail.
Mapping high-dimensional inputs into compact, learnable vector spaces.
Identity shortcuts and gradient flow — why ResNets train deep networks.
Pooling sequence representations into single, decision-ready vectors.
Multi-task heads: separate output paths for action, value and risk.
Composite objectives and risk-aware losses.
Gradient-based optimisers — SGD, Adam, AdamW — and learning-rate schedules.
Composite objectives: return, drawdown and Sharpe combined into a single loss.
Encoding risk constraints directly into the training signal.
A penalty term that rewards trading only when liquidity is sufficient.
Stable, scalable, debuggable training.
Classical supervised pre-training before reinforcement-learning fine-tuning.
When deep networks stop learning — and how residuals and normalisation fix it.
Bounded updates to prevent explosive gradients during training.
Standardising inputs for stable, fast convergence.
Training across multiple workers — synchronous and asynchronous all-reduce.
Filling missing market data with statistically consistent values.
Tools for inspecting what a trading model actually learned.
How error is measured shapes what is learned.
An L1/L2 hybrid loss — robust to outliers without losing precision near zero.
A loss that distinguishes direction errors from magnitude errors.
A loss that scales penalty with the magnitude of the prediction error.