Scholarlink

<p>Forecasting stock prices is complicate due the financial market’s nature, which includes volatility, non-stationarity, and reliance on information for decision-making. Stock prices respond to a multitude of information sources including technical indicators, historical price movements, news articles, and investor sentiments. The recent introduction of FinBERT-based transformer models has also improved prediction accuracy; however, the majority of existing methodologies utilize static train-test splits, limited feature selection, weak cross-modal fusion, and do not take into account the impact of changing market regime dynamics, resulting in lower-than-expected robustness when applied to changing market conditions. This paper presents a new hybrid forecasting framework that enhances the prediction ability using walk-forward training; performs cross-modal feature selection; includes awareness of market regime; and reduces prediction bias with the use of a meta-classifier for the final prediction output. The framework developed in this study combines the use of sentiment features from FinBERT with price-based technical indicators to conduct stock price forecasts. Hidden Markov Models are used to identify market regimes, and the temporal dependencies is learned through the application of a CNN-Transformer architecture. Mutual information based feature selection is used to reduce the redundancy of features, and the final stock price forecast will be produced through the application of a meta-classifier. A walk-forward evaluation is used in order to ensure robustness to non-stationary stock price movement data. The average accuracy for experimental results on ten major U.S. stocks is 70%, F1-macro is 68.0%, MCC is 61.6% and AUC is 76.5%. Backtests incorporating realistic transaction costs produced a win rate greater than 60.07% and average returns of 30.82% which exceeded the performance of a pure deep learning-only baselines by a factor of (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(p &lt; 0.05\)</EquationSource> </InlineEquation>). Thus, this work presents a robust, regime-aware and leakage resistant stock forecasting framework which displays superior predictive and trading capabilities when compared to other methods simulating actual market conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A walk-forward multi-model approach to stock trading: leveraging sentiment, trend indicators, and transformer-based deep learning for next-day price forecasting

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Shahabaj Khan,
  • Anil Kumar Kushwah,
  • Jagdish Chakole

摘要

Forecasting stock prices is complicate due the financial market’s nature, which includes volatility, non-stationarity, and reliance on information for decision-making. Stock prices respond to a multitude of information sources including technical indicators, historical price movements, news articles, and investor sentiments. The recent introduction of FinBERT-based transformer models has also improved prediction accuracy; however, the majority of existing methodologies utilize static train-test splits, limited feature selection, weak cross-modal fusion, and do not take into account the impact of changing market regime dynamics, resulting in lower-than-expected robustness when applied to changing market conditions. This paper presents a new hybrid forecasting framework that enhances the prediction ability using walk-forward training; performs cross-modal feature selection; includes awareness of market regime; and reduces prediction bias with the use of a meta-classifier for the final prediction output. The framework developed in this study combines the use of sentiment features from FinBERT with price-based technical indicators to conduct stock price forecasts. Hidden Markov Models are used to identify market regimes, and the temporal dependencies is learned through the application of a CNN-Transformer architecture. Mutual information based feature selection is used to reduce the redundancy of features, and the final stock price forecast will be produced through the application of a meta-classifier. A walk-forward evaluation is used in order to ensure robustness to non-stationary stock price movement data. The average accuracy for experimental results on ten major U.S. stocks is 70%, F1-macro is 68.0%, MCC is 61.6% and AUC is 76.5%. Backtests incorporating realistic transaction costs produced a win rate greater than 60.07% and average returns of 30.82% which exceeded the performance of a pure deep learning-only baselines by a factor of ( \(p < 0.05\) ). Thus, this work presents a robust, regime-aware and leakage resistant stock forecasting framework which displays superior predictive and trading capabilities when compared to other methods simulating actual market conditions.