News → Markets (Macro USA) – Starter Project

Inputs (place them here):

data/raw/dataset.csv (your GDELT filtered news export)
data/raw/market_data.csv (your yfinance export)

Pipeline:

Build merged ML dataset (daily features + next-day returns).
Train a simple neural net baseline (PyTorch) you can tune.

Install

pip install -r requirements.txt

Build dataset

python -m news_impact.prepare_dataset --news data/raw/dataset.csv --market data/raw/market_data.csv --out data/processed/ml_dataset.parquet

Train baseline NN

python -m news_impact.train --data data/processed/ml_dataset.parquet --targets UUP TLT GLD BTC-USD XLK XLF XLE XLV XLI XLY XLP XLB XLU XLRE XLC

Outputs are saved to runs/<timestamp>/.

Extended training/evaluation pipeline

The final dataset should be at:

data/processed/ml_dataset.parquet

Train one model (CLI)

python -m news_impact.train --data data/processed/ml_dataset.parquet --model baseline_zero --outdir models/baseline_zero
python -m news_impact.train --data data/processed/ml_dataset.parquet --model baseline_last --outdir models/baseline_last
python -m news_impact.train --data data/processed/ml_dataset.parquet --model knn --outdir models/knn
python -m news_impact.train --data data/processed/ml_dataset.parquet --model ridge --outdir models/ridge
python -m news_impact.train --data data/processed/ml_dataset.parquet --model rf --outdir models/rf
python -m news_impact.train --data data/processed/ml_dataset.parquet --model mlp --outdir models/mlp

Evaluate a trained model

python -m news_impact.evaluate --data data/processed/ml_dataset.parquet --modeldir models/knn --outdir reports

Full runner (train + eval + leaderboard)

# quick (default)
python -m news_impact.run_all --data data/processed/ml_dataset.parquet

# full
python -m news_impact.run_all --data data/processed/ml_dataset.parquet --mode full

Prediction (inference)

# explicit model
python -m news_impact.predict --data data/processed/ml_dataset.parquet --modeldir models/rf --out reports/predictions/preds.csv

# best global model by metric
python -m news_impact.predict --data data/processed/ml_dataset.parquet --best --best-by directional_accuracy --best-scope global --out reports/predictions/preds.csv

# best model per target (mixed models)
python -m news_impact.predict --data data/processed/ml_dataset.parquet --best --best-by directional_accuracy --best-scope per-target --out reports/predictions/preds.csv

# simple backtest (long-flat / long-short)
python -m news_impact.backtest --reports-dir reports --cost-bps 1.0

# backtest with confidence filter + sweep
python -m news_impact.backtest --reports-dir reports --cost-bps 1.0 --thresholds 0,0.0005,0.001,0.002,0.003,0.005

# final operational selection (global / per-target)
python -m news_impact.select_setup --reports-dir reports --objective sharpe --cost-bps 1 --scope global
python -m news_impact.select_setup --reports-dir reports --objective sharpe --cost-bps 1 --scope per-target

# inference using final selection
python -m news_impact.predict --data data/processed/ml_dataset.parquet --selection-file reports/selection/final_selection_per_target.csv --out reports/predictions/preds.csv

Artifacts:

Models: models/<model>/
Metrics: reports/metrics/
Figures: reports/plots/<model>/
Summary: reports/summary.md
Predictions: reports/predictions/preds.csv

The pipeline includes:

multi-target regression (15 y_* targets)
time-ordered split (no shuffle) + configurable expanding walk-forward
per-target and average metrics: MAE, RMSE, R², directional accuracy / hit rate, Pearson, Spearman
required baselines: baseline_zero, baseline_last
kNN “similar days” baseline

pliegopera/ann_news_impact

About

Resources

Stars

Watchers

Forks

Releases

Languages