Instrumental analysis and forecasting of financial market behavior using NLP and LSTM
Loading...
Date
2026
Authors
Hrynchuk, Tymofii
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Long Short-Term Memory (LSTM) networks remain a widely used baseline for financial time-series forecasting, but most published studies focus on a single prediction target with a fixed feature set. This makes it difficult to answer a practical question: which types of additional information actually help, and for which prediction tasks? This research investigates the contribution of two feature modalities – NLP-derived sentiment and numerical technical analysis (TA) patterns – to four specialised LSTM prediction tasks on 4-hour BTCUSDT data from January 2022 to January 2026. The four tasks are: next-candle return regression, three-class directional classification (DOWN / NEUTRAL / UP), ZigZag reversal detection, and pump/dump event classification. Each task is handled by an independent LSTM model, allowing taskspecific loss functions and Optuna hyperparameter searches. A controlled ablation study runs each model under four feature configurations: Experiment A (53 baseline features), Experiment B (+2 NLP features, 55 total), Experiment C (+14 TA pattern features, 67 total), and Experiment D (all features combined, 69 total). The NLP pipeline processes over 3,5M deduplicated texts from seven sources through FinBERT to produce per-window sentiment scores. Results show that feature relevance is strongly task-dependent. NLP sentiment provides the largest marginal improvement for return prediction and directional accuracy. TA patterns provide the largest improvement for reversal detection (AUC rising from 0.668 to 0.703). Pump and dump detection benefits substantially only when both modalities are combined (Experiment D raises accuracy from 69.2% to 81.5% and event F1 from 26.8% to 31.3%).
Description
Keywords
Long Short-Term Memory (LSTM), financial time-series forecasting, FinBERT sentiment analysis, technical analysis patterns, BTCUSDT, ablation study, multi-task prediction, bachelor`s thesis