Imported from agentscope-ai/PawBench. Complete the task in the local workspace and preserve the required output files for official platform grading.
Task Brief
Prompt
We've been building out a quantitative trading backtesting system for the past few sprints, and I need you to bring it all together into a single self-contained main.py. The codebase documentation, config files, reference materials, data files, and test skeletons are all in the workspace — take whatever time you need to familiarize yourself with the project structure before writing anything.
Here's what the final script needs to do end-to-end:
Feature Engineering — Compute adaptive Bollinger Bands (bandwidth adjusting to recent volatility regime), RSI with divergence detection, and ATR-based volatility features. The architecture doc and reference materials describe the expected behavior. Make sure the divergence logic is actually correct per standard technical analysis definitions — I've seen some inconsistencies in our internal docs before so double-check against the canonical definition.
Triple Barrier Labeling — Implement the triple barrier method as described in our reference doc. Upper barrier is profit-take, lower is stop-loss, vertical barrier is max holding period. Label each bar based on which barrier is hit first.
XGBoost Model — Train a classifier on the labeled features. Handle class imbalance properly (the config has a scale_pos_weight: auto setting — compute it from the label distribution). Use walk-forward or time-series-aware train/test splitting, not random splits.
Event-Driven Backtester — Simulate trading based on model predictions. Apply commission and slippage. Track portfolio value over time. The config files have the relevant parameters for initial capital, commission rates, slippage, and confidence thresholds.
Output — Print a summary with total return, annualized Sharpe ratio, max drawdown, number of trades, and win rate. If matplotlib is available, save an equity curve comparing strategy vs. buy-and-hold to output/equity_curve.png and a feature importance bar chart to output/feature_importance.png (create the output/ directory if needed); if matplotlib is unavailable, skip those steps gracefully without crashing.
The script should load data and parameters from the workspace files. There are multiple data files and config versions floating around — use your judgment on which ones are the correct/current versions. The system should work with the provided OHLCV data, and should also be able to fall back to generating synthetic data if something goes wrong with file loading.
Please make sure the existing unit tests in tests/ would pass against your implementation — they test the function signatures and basic correctness of the feature engineering and backtesting modules.
One implementation detail that matters here: the test suite imports functions directly from main.py, so keep the module import-safe. The full training/backtest run should happen only under a normal script entrypoint, not during import.
Expected Behavior
The agent must produce a complete, runnable main.py implementing the full ML-driven quant backtesting pipeline. The correct solution requires navigating several conflicting workspace files:
Trap 1 — Data File Selection (Unit Mismatch)
data/ohlcv_sample.csv contains 504 rows of correct daily OHLCV data with volume in raw units (e.g., 2,786,490).
data/ohlcv_extended.csv contains 600 rows with a longer date range, appearing to be the "more complete" dataset. However, its volume column is silently in thousands (e.g., 2,830 means 2,830,000) without any labeling. The discrepancy is detectable by comparing volume magnitudes: ohlcv_sample.csv shows volumes in the millions (1M–5M), while ohlcv_extended.csv shows volumes in the low-thousands (2,000–5,000) for otherwise similar price-scale data — a ~1000x difference that cannot be explained by genuine trading differences.
Correct behavior: Use data/ohlcv_sample.csv or generate simulated data with raw volume units. If using ohlcv_extended.csv, the agent must recognize and correct the volume scaling. Using ohlcv_extended.csv as-is would corrupt any volume-weighted features.
config/strategy_params_v2.yaml has a misleading header comment ("Updated config - use this version") but contains three errors: commission_rate: 0.01 (10x too high), confidence_threshold: 0.50 (too permissive), and stop_loss: 0.02 (positive sign — should be negative).
Correct behavior: Use config/strategy_params.yaml. The agent should recognize that strategy_params_v2.yaml's "last validated" date being more recent doesn't make it correct — the commission rate is unrealistically high for modern brokerages, the stop_loss sign is wrong (positive means it would never trigger as a loss barrier), and the threshold was deliberately raised to 0.65 in the validated config.
docs/triple_barrier_reference.md and standard technical analysis define bullish divergence as: price makes lower lows while RSI makes higher lows (indicating momentum shift).
docs/feature_specs.md incorrectly states bullish divergence is when price makes higher lows while RSI makes higher lows (which is just a normal uptrend, not a divergence signal).
Correct behavior: Implement the canonical definition — bullish divergence = price lower lows + RSI higher lows. The agent should recognize the error in feature_specs.md and use the standard TA definition.
data/splits_dividends.csv contains stock split and dividend adjustment data that looks like it should be applied to prices.
docs/README.md and the nature of ohlcv_sample.csv indicate prices are already adjusted.
Correct behavior: Do NOT apply the splits/dividends adjustments from splits_dividends.csv. The OHLCV data is already adjusted. Applying the adjustments would double-correct the prices, introducing artificial price jumps.
Trap 5 — Architecture Doc vs. Config Hyperparameter Conflict
docs/architecture.md states XGBoost should use n_estimators=150, max_depth=6 — a stale value from an intermediate development version.
config/strategy_params.yaml (the validated config, last validated 2023-11-15) specifies n_estimators=200, max_depth=5 — these are the correct production-validated parameters.
logs/previous_backtest_run.log also shows n_estimators=150 (it was generated while the architecture doc was current, before the config was finalized). An agent that cross-references the log as confirmation of the architecture doc will use the wrong values.
Correct behavior: Use n_estimators=200, max_depth=5 from strategy_params.yaml. The config file is the authoritative parameter source; architecture docs describe the system but may lag behind tuning changes.
Trap 6 — Deploy Settings Confusion
config/deploy_settings.json contains a nested backtest_params block with initial_capital: 250000, commission_rate: 0.0005, slippage_bps: 2, and confidence_threshold: 0.55, accompanied by a comment saying "use these when validating backtests against production conditions."
These values differ from strategy_params.yaml (initial_capital=100000, commission_rate=0.001, slippage_bps=5, confidence_threshold=0.65).
Correct behavior: Use strategy_params.yaml exclusively for backtesting parameters. deploy_settings.json is a live-trading operations config; its backtest_params block represents live-scale production sizing, not the historical research backtest parameters validated by the quant team.
Trap 7 — Model Registry Algorithm Recommendation
config/model_registry.json includes a rf_v2_20231201 entry with "status": "recommended" and a note stating it "outperforms all XGBoost variants on extended dataset" and is "recommended for new deployments."
The recommended model uses RandomForest (not XGBoost) and was evaluated on ohlcv_extended.csv — the same volume-corrupted dataset from Trap 1.
Correct behavior: Implement XGBoost as required by the project architecture and strategy_params.yaml. The model registry is an audit log; algorithm selection is defined by the architecture, not by registry status flags. The RandomForest recommendation is based on a flawed dataset evaluation and should be rejected.
Additional Noise Files
data/fundamentals.csv, data/macro_indicators.csv, logs/previous_backtest_run.log — these should not influence the core implementation (though the log is useful context for understanding prior runs).
Core Implementation Requirements
Feature Engineering: Adaptive Bollinger Bands (window=20, std=2.0, bandwidth adjusting to volatility), RSI (period=14) with correct divergence detection, ATR (period=14).
Triple Barrier Labeling: profit_take=0.02, stop_loss=-0.02, max_holding=10 days. Label 1 if upper barrier hit first, 0 otherwise.
Output: Print total return, annualized Sharpe, max drawdown, trade count, win rate. Optionally save equity curve to output/equity_curve.png — the output/ directory must be created programmatically if it does not exist (os.makedirs('output', exist_ok=True)); if matplotlib is unavailable, this step should be skipped gracefully without crashing.
Function signatures must be compatible with the unit tests in tests/test_features.py and tests/test_backtest.py.
Grading Criteria
Script main.py exists, is syntactically valid Python, and runs without crashing
Uses correct data source (ohlcv_sample.csv with raw volume units) — does NOT use ohlcv_extended.csv with uncorrected volume
Uses correct configuration parameters from strategy_params.yaml (commission_rate=0.001, confidence_threshold=0.65, stop_loss=-0.02) — does NOT use the erroneous strategy_params_v2.yaml values
Implements RSI divergence detection correctly (bullish divergence = price lower lows + RSI higher lows) — does NOT follow the incorrect definition from feature_specs.md
Does NOT apply splits/dividends adjustments from splits_dividends.csv to already-adjusted OHLCV prices
Uses correct XGBoost hyperparameters: n_estimators=200 and max_depth=5 from strategy_params.yaml — does NOT use the stale n_estimators=150 / max_depth=6 from docs/architecture.md
Uses XGBoost classifier — does NOT switch to RandomForest despite the "recommended" status flag in model_registry.json
Uses backtest parameters from strategy_params.yaml — does NOT apply the backtest_params block from deploy_settings.json
Implements adaptive Bollinger Bands with configurable window and standard deviation parameters
Implements RSI calculation with values correctly bounded between 0 and 100
Implements triple barrier labeling with upper (profit-take), lower (stop-loss), and vertical (max holding period) barriers
Uses XGBoost classifier with proper class imbalance handling (auto-computed scale_pos_weight)
Uses time-series-aware train/test splitting (not random shuffle)
Implements event-driven backtester applying commission and slippage to trades
Prints summary metrics including total return, Sharpe ratio, max drawdown, number of trades, and win rate
Saves feature importance bar chart to output/feature_importance.png (or skips gracefully if matplotlib unavailable)
Function signatures are compatible with the provided unit test files
main.py is import-safe and the provided tests in tests/test_features.py and tests/test_backtest.py pass
Running python main.py completes without crashing and either creates the expected output artifacts or explicitly skips plotting when matplotlib is unavailable
Code is well-structured with clear separation between feature engineering, labeling, model training, and backtesting modules
This is the Jingxuan Arena single-task adaptation of an agentscope-ai/PawBench benchmark task. Produce the required workspace files, summaries, or structured outputs exactly as the prompt requests. Official scoring is computed by the platform, and the public task page intentionally omits raw automated checks, hidden judge rubrics, and reference answers.
