anonymous-submission-678/backtrader-mcq-base-pool-all-strategies

Backtrader MCQ Benchmark

This dataset contains multiple-choice questions for evaluating whether a model can reason about trading-strategy behavior using the Backtrader backtesting framework. Each question provides a complete backtest configuration and asks for a single answer choice in the format <<< X >>>, where X is one of A, B, C, or D.

The primary evaluation file used in the paper is:

backtrader_mcq_balanced_30_all_strategies.jsonl

The larger supporting pool is:

backtrader_mcq_base_pool_all_strategies.jsonl

Dataset Files

File	Examples	Role
backtrader_mcq_balanced_30_all_strategies.jsonl	30	Balanced evaluation subset used in the paper
backtrader_mcq_base_pool_all_strategies.jsonl	160	Larger base pool from which evaluation questions can be inspected or sampled

The balanced evaluation subset contains:

• 10 easy, 10 medium, and 10 hard questions.
• 6 questions for each strategy.
• Questions over AAPL from 2020-01-01 to 2024-01-01.

The base pool contains 160 questions across the same benchmark setting, with a broader distribution of question types and difficulty levels.

Task

Given a question prompt, a model must select the correct multiple-choice answer and respond exactly as:

<<< A >>>

or the corresponding option letter. The ground-truth answer is stored in the answer field.

Many questions cannot be answered from surface text alone. They require computing or verifying values from the specified Backtrader run, including portfolio value, broker cash, positions, orders, trades, drawdowns, moving averages, and strategy comparisons.

Strategies

The dataset covers five Backtrader strategies:

• sma_crossover
• rolling_window_mean
• exponential_weighted_moving_average
• rsi_strategy
• macd_crossover

Each row includes the exact strategy_name, strategy_params, and full config needed to reproduce the corresponding backtest.

Schema

Each JSONL row has the following fields:

Field	Description
question	Full multiple-choice prompt, including the backtest configuration and answer choices
answer	Correct option letter, one of A, B, C, or D
benchmark	Benchmark identifier; currently backtrader
author	Internal generation label
type	Question category, such as portfolio_value, max_drawdown, first_buy, or roi_percentage
difficulty	Difficulty label: easy, medium, or hard
strategy_name	Backtrader strategy used for the question
strategy_params	Strategy parameters used for the question
config	Full backtest configuration, including symbol, date range, cash, commission, stake, and strategy settings

Example record shape:

{
  "question": "...",
  "answer": "C",
  "benchmark": "backtrader",
  "author": "MCQ_test",
  "type": "first_close",
  "difficulty": "easy",
  "strategy_name": "sma_crossover",
  "strategy_params": {
    "pfast": 10,
    "pslow": 30,
    "target_percent": 1.0
  },
  "config": {
    "symbol": "AAPL",
    "start_date": "2020-01-01",
    "end_date": "2024-01-01",
    "initial_cash": 50000.0,
    "commission_rate": 0.001,
    "stake": 1000,
    "strategy_name": "sma_crossover",
    "strategy_params": {
      "pfast": 10,
      "pslow": 30,
      "target_percent": 1.0
    }
  }
}

Data Generation

Questions are generated from deterministic Backtrader runs over historical AAPL market data loaded with yfinance using auto_adjust=False. Each run uses the row-level configuration embedded in the JSONL record.

The standard configuration used by the released questions is:

• Symbol: AAPL
• Date range: 2020-01-01 to 2024-01-01
• Initial cash: 50000.0
• Commission rate: 0.001
• Stake cap: 1000

Backtrader is used to compute the daily records and trading events needed for each question. These include OHLCV-derived values, moving averages, broker cash, portfolio value, position size, order executions, commissions, closed-trade PnL, and derived metrics such as ROI, drawdown, profit factor, and Calmar-style ratio.

Evaluation

For the paper setting, evaluate on:

backtrader_mcq_balanced_30_all_strategies.jsonl

Recommended scoring:

1. Prompt the model with the question field.
2. Parse the predicted answer choice from the response.
3. Compare the predicted choice against the answer field.
4. Report exact-match multiple-choice accuracy.

The answer should be treated as the option letter, not the full answer text.

Intended Use

This dataset is intended for evaluating tool-use, programmatic reasoning, and quantitative reasoning in the context of trading backtests. It is especially suited for testing whether models can correctly use Backtrader or equivalent computations to answer finance-oriented multiple-choice questions.

This dataset is not intended to provide financial advice, trading recommendations, or claims about future market behavior.

Limitations

• The released questions use one ticker, AAPL, over one historical date range.
• Correct answers depend on the exact Backtrader implementation, strategy definitions, commission model, position sizing logic, and market data version.
• Historical market data can change across vendors or download dates due to adjustments, corrections, or API behavior.
• The dataset is small and should be interpreted as a focused benchmark rather than a broad measure of financial expertise.
• Multiple-choice formatting may allow some models to exploit answer-choice artifacts; tool-based verification is recommended.

NeurIPS Review Notes

This repository is prepared for anonymous review. The dataset should remain accessible to reviewers, ACs, and SACs at submission time according to the NeurIPS Evaluations & Datasets Track requirements. The repository metadata, README, and file contents should avoid revealing author identity during the review period.

For NeurIPS Evaluations & Datasets submissions, the Hugging Face dataset page can generate core Croissant metadata automatically. Please ensure the required Responsible AI fields are completed before submission if required by the track instructions.

License

The license is marked as other during anonymous review. Final public release terms should be confirmed before camera-ready release.