Findeep is a robust Financial Forecasting Pipeline that predicts next-quarter corporate revenues using fundamental financial data from SEC filings (10-K and 10-Q).

It features a strict Backtesting Framework designed to simulate real-world trading/forecasting conditions by preventing lookahead bias and data leakage.

• Automated Data Ingestion: Custom scraping pipelines for SEC EDGAR that fetch, parse, and structure XBRL/XML data from 10-Q (Quarterly) and 10-K (Annual) reports.
• Leakage-Free Backtest Engine: A specialized validation framework (run_proper_backtest.py) that strictly segregates training and testing data by time. It ensures that predictions for Quarter $T$ use only information available before $T$.
• Advanced Feature Engineering: Generates time-series features such as quarter-over-quarter growth rates, lagged indicators ($t-1, t-2$), and rolling means.
• Machine Learning: Uses XGBoost (Gradient Boosting) for regression forecasting, tuned for small-sample financial time-series.

• run_proper_backtest.py: Main Entry Point. Orchestrates the entire backtesting workflow (Load Data -> Train -> Predict -> Evaluate).
• tesla_10q_pipeline.py: Data pipeline for Tesla (TSLA). Fetches 10-Q/10-K filings and extracts financial tags (Revenue, Assets, etc.).
• JP_10q_pipeline_full.py: Data pipeline for JP Morgan (JPM).
• filter_and_combine_excel.py: Utility to merge individual company datasets into a unified format for training.
• src/: Core logic library.
  • backtest_utils.py: Contains the assert_no_leakage logic and test set construction.
  • supervised_prep.py: Feature engineering logic (lag generation).
  • train_next_quarter.py: XGBoost model training wrapper.

Ensure you have Python 3.10+ installed.

pip install -r requirements.txt

Fetch the latest financial data directly from the SEC:

# Fetch Tesla Data (10-Q & 10-K)
python tesla_10q_pipeline.py

# Fetch JP Morgan Data (10-Q & 10-K)
python JP_10q_pipeline_full.py

Combine the raw excel files into a training dataset:

python filter_and_combine_excel.py

Execute the full forecasting pipeline:

python run_proper_backtest.py

This will:

1. Load the combined data.
2. Train an XGBoost model on historical data (e.g., up to 2024-03-31).
3. Predict the specific target quarter (e.g., 2024-06-30).
4. Save results to artifacts/backtest_predictions.parquet.
5. Print evaluation metrics (MAE, RMSE).

Unlike standard cross-validation, financial time-series require strict temporal ordering. This project implements a Walk-Forward Validation approach.

• Training: Uses all available history up to train_end_date.
• Testing: Predicts exactly one step ahead (test_date).
• Safety: The src/backtest_utils.py module enforces that feature generation for the test set never accesses future rows, preventing the common "lookahead bias" error in financial ML.