Kernel: from a research platform to an autonomous trading agent

The challenge

A single chart or several profitable trades cannot validate a strategy. A repeatable environment was needed to automate data collection, prediction, simulation, cost modelling and comparable evaluation.

Data volume and velocity were another challenge. During active periods the platform processes dozens of market transactions per second, which required deliberate retention, aggregation and compression policies.

The solution

Kernel provides one domain model for live data, historical replay, virtual accounts, strategies and machine-learning models.

It automatically runs simulations for many parameter sets and compares profitability, stability, risk and cost impact. Every result is connected to the input data, strategy version, model version and calculated costs, making experiments reproducible.

Direction: an autonomous trading agent

The research platform is a controlled stage rather than the project’s final destination. The target is an agent operating autonomously or semi-autonomously, depending on risk limits and the decisions reserved for human approval.

The agent is intended to monitor markets continuously, generate predictions, select a strategy and its parameters, simulate decision outcomes and manage exposure. In semi-autonomous mode, the operator receives rationale, context and a risk assessment before approving an action.

Kernel records simulated decisions and their later outcomes. Successive cycles use this material to train and evaluate new model versions from the beginning under the same data and criteria.

Key capabilities

• real-time and historical market data
• dozens of market transactions processed per second
• automated testing of strategies and parameter variants
• backtesting, replay and virtual accounts
• fees, spread, slippage and latency simulation
• result ranking and exposure control
• model training, validation, versioning and metrics
• a complete audit trail of decisions, data, model versions and outcomes

Machine-learning models

The current version uses proprietary models built with XGBoost and LightGBM. They return direction and confidence estimates that feed a controlled strategy engine incorporating risk, cost and exposure rules.

At this stage the platform does not autonomously trade real capital. Simulated decisions and their outcomes provide training material for subsequent model candidates.

Data-layer optimisation

TimescaleDB storage was reduced from approximately 34 GB to 9.5 GB, a reduction of around 72%, while retaining the key information required for further experiments.

Research and business outcome

The result is an operational environment automating the research loop from continuous data collection through prediction, simulation and model training to comparable evaluation.

Most importantly, it provides an auditable foundation for a future trading agent: one process joining market observation, a decision, simulation of its effect, evaluation and retraining.