news-sentiment-research

Run the backtest scripts inside the quantmind Docker container, which load enriched sentiment from PostgreSQL, timestamps from Huntly SQLite, and forward-adjusted daily klines from QuantDB parquet. The optimized script supports toggles for source filtering, time filtering, multi-article confirmation, and momentum filters.

Three sources are joined: Huntly SQLite for article timestamps and source names, a PostgreSQL news_article_enrichment table for tickers, sentiment labels, scores and event tags, and QuantDB parquet files for daily prices and the trading calendar.

Testing showed fixed stop-loss exits had 0% win rate because news-driven moves typically dip before rallying. The strategy instead uses a trailing take-profit activated above 15% gain with a 5% drawdown exit, which was the only profitable exit mechanism.

No. The scripts import pandas, numpy, psycopg2, sqlalchemy and internal backend packages that only exist in the quantmind container. They must be executed via docker exec, while pure standard-library scripts may run locally.

Bullish news effects decay within about 15 trading days and are strongest in T+1 to T+3. Some sources are reverse indicators, ChiNext bearish signals were ineffective, and results depend on the data snapshot since the enrichment table updates continuously.