Accounting for Time Decay in Sentiment-Based Trading Signals - A Comparative Study in Algorithmic Trading

Sammanfattning

Algorithmic trading has become increasingly data-driven, and financial news and market prices are often used together to generate trading signals. Since the relevance of financial news may decrease over time, it is important to investigate whether sentiment signals should be weighted by recency rather than treated equally within a fixed time window. This study examines whether modelling time decay in sentiment-based trading signals improves the performance of an algorithmic trading strategy. To address this question, an experiment was conducted in which two otherwise identical strategy variants were compared: one using decay-weighted aggregation of news sentiment over time and one using equal-weight aggregation within the same time window. A trading system was implemented that combined sentiment-based information from financial news headlines with price-based prediction from historical market data. The two strategy variants were evaluated under identical conditions through historical backtests and live forward testing. Performance was assessed using win rate, expectancy in risk units, profit factor, Sharpe ratio, and maximum drawdown. The results indicate that the effect of time decay was mixed rather than uniform. In the backtests, the decay-weighted strategy performed better for data from 2022 and showed a small overall advantage for data from 2023, while the non-decayed strategy performed better on most metrics for data from 2024. The forward test likewise did not show a consistent overall advantage for either variant. Overall, explicit time-decay modelling influenced strategy performance, but did not lead to a clear and consistent improvement across all evaluated settings.