AAPL-QQQ Lead-Lag

You have 8 hours of wall-clock time for this task. Your machine has 16 CPUs, 110000 MiB of RAM and no GPU. When the time is up, you are stopped and your work is graded as it stands.

We run a reference AAPL strategy that trades a QQQ cross-asset lead-lag signal — the index-implied move of AAPL, gated to fire only when AAPL lags its index — off the IEX DEEP+ DPLS event tape.

Your mandate is to produce an updated strategy that beats it on hidden out-of-sample (OOS) data. Form a view on where the edge is, design experiments that could prove you wrong, read what comes back honestly, and keep only what survives. We care about edge that holds up day after day, not a single good backtest.

Environment

You have the full research loop here: inspect the real market data, define trading strategies across a range of horizons, and simulate how they perform.

• /src/env/trading_system/: Trading system code. Read the README and source files to understand the interface you must implement, the available features, and the simulator.
• /src/env/artifacts/: Research artifacts and reports that may help you understand the reference strategy and feature ideas.
• /data/: Market data for research.

All required code and data is in /src/, /data/, and /workspace/.

Baseline metrics

You may first run python /src/env/trading_system/research/tools/build_baseline_metrics.py to establish the baseline train and validation metrics.

Save current-best strategies

Use python /src/env/trading_system/engine/scripts/save_strategy.py to checkpoint the best strategy you have found so far. If there is no saved strategy yet, save once you have validation evidence that the current implementation is better than the reference lead-lag strategy. After you have saved a strategy, save again only when you believe the current implementation is better than your existing saved best. Every saved strategy is retained, and you are evaluated based on your last saved strategy snapshot under /save. The model from your last checkpoint will not be retrained for evaluation, and later edits to /src/env/trading_system/research/ do not affect grading unless you run save_strategy.py again.

Your final saved strategy will be evaluated on hidden OOS data against the reference lead-lag strategy by loading your trading strategy (model, training configuration features, taking logic, etc.) and simulating on hidden OOS data. Your goal is to have consistent daily PnL outperformance against the reference strategy.

Submission requirements

The final strategy must avoid lookahead bias: strategy logic and learned parameters may use only information available at each decision point. Any learned parameters used in the final strategy must be embedded in code or fit causally from past data only.