Architecture¶

Overview¶

tradedesk is an event-driven trading framework designed to run strategies across both backtesting and live broker environments.
The codebase is organized into distinct domains:
tradedesk.marketdata: ingestion and normalization of price data and events.
tradedesk.data_sources: free, no-auth macro datasets (FRED rates and VIX, ECB €STR / yield curve / Euribor, CFTC COT positioning) materialized to a Parquet lake and loaded via a single load_macro_series / load_macro_frame access pattern.
tradedesk.strategy: base classes and helpers for writing event-driven strategies.
tradedesk.portfolio: portfolio management, risk budgeting, and multi-instrument coordination.
tradedesk.execution: execution adapters (IG for live trading, Dukascopy paths for backtests).
tradedesk.recording: trade lifecycle tracking, metrics, and reporting.
tradedesk.research: research helpers such as walk-forward tooling, correlation gates, and public configuration templates.
tradedesk.ml: machine-learning building blocks for feature engineering, labels, walk-forward CV, and model wrappers.

Market data events flow into strategies via the event system.
Strategies emit execution requests in response to data events.
Portfolio layer applies risk budgets and distributes them to active sleeves/instruments.
Execution layer applies pre-flight gates (for example spread limits or portfolio-level order gates), then places orders and emits order-completion plus position-lifecycle events.
Recording layer captures lifecycle events and metrics for reporting.

Live trading uses the IG module (IGClient, IGAuthManager, etc.).
Backtesting uses the Dukascopy-based data path, converting Dukascopy data to the internal candle/tick format.

Pure event-driven components: strategies react to events, not to direct broker state.
Separate concerns: data handling, strategy logic, risk management, and execution are decoupled to simplify testing and maintenance.
Deterministic backtesting: same strategy code runs in both backtest and live modes to ensure reproducibility.
Live parity for recording: when a client does not publish its own position-open events, the execution layer emits PositionOpenedEvent after a confirmed opening fill so recording and downstream observers see the same lifecycle boundary in backtest and IG-backed sessions.

Licensed under the Apache License, Version 2.0. See: https://www.apache.org/licenses/LICENSE-2.0