How AI can transform financial forecasting and decision-making in fintech environments

A new study provides one of the most rigorous comparisons of advanced deep learning models used for long-term stock market prediction. Titled "AI-Driven Intelligent Financial Forecasting: A Comparative Study of Advanced Deep Learning Models for Long-Term Stock Market Prediction", the research was published in Machine Learning and Knowledge Extraction (MAKE).

The paper evaluates 10 cutting-edge models, including transformer-based architectures, using historical data from the S&P 500, NASDAQ, and Hang Seng indices. By combining traditional predictive accuracy metrics with financial performance indicators such as returns, volatility, drawdown, and Sharpe ratios, the study delivers crucial insights into which AI models can truly support smart, risk-aware investment strategies over extended time horizons.

Which AI models perform best across different forecasting horizons?

The study benchmarked the effectiveness of ten deep learning models across four forecasting horizons: 96, 192, 336, and 720 days. The analysis spanned both transformer-based and non-transformer models, including PatchTST, Crossformer, Autoformer, Informer, TimeNet, and DLinear.

The findings show that PatchTST consistently delivered the best predictive accuracy across short- and mid-term horizons for the S&P 500 and NASDAQ datasets. The model’s patch-based approach allowed it to capture local temporal patterns effectively. However, in extended forecasting scenarios, simpler models like DLinear demonstrated surprising resilience and outperformed more complex architectures like Informer or Crossformer for the 720-day horizon.

On the performance front, Crossformer emerged as a high-return model, offering gains of over 113% on the 720-day forecast, albeit with higher volatility and significant drawdowns. In contrast, models like Autoformer and Non-stationary Transformer delivered more conservative but stable performance metrics, making them more suitable for risk-averse strategies.

These variations in results suggest that no single model dominates across all use cases. Model selection should be tightly aligned with the investment time horizon and risk profile.

How were the models evaluated beyond accuracy?

Unlike most prior studies that emphasize accuracy metrics alone, this research implemented a dual-metric evaluation framework, combining Mean Absolute Error (MAE) and Mean Squared Error (MSE) with four essential financial indicators:

• Return – Measured as the cumulative profit or loss over the forecasting period.
• Volatility – Standard deviation of returns, used to assess stability.
• Maximum Drawdown – Peak-to-trough loss, indicating exposure to sharp declines.
• Sharpe Ratio – A risk-adjusted performance metric.

The models were tested using a simplistic long-only trading strategy based on whether the predicted price exceeded the current closing price. While this rule may not reflect real-world complexity, it isolates the model’s directional prediction ability.

Among the top financial performers:

• Transformer reached a return of 139.63% over 720 days but with the highest volatility (72.81%).
• Informer also performed well financially but suffered from inconsistent drawdowns.
• PatchTST, despite its early dominance, exhibited weaker long-horizon returns and higher drawdowns, suggesting it may be best suited for tactical short-term strategies.

Notably, statistical validation using the Mann–Whitney U test confirmed that many of these performance differences were significant, especially across changing time horizons.

What are the practical implications for investors and fintech?

Forecasting horizon matters significantly. While some models shine in short-term scenarios, they may falter over longer stretches. Secondly, model complexity does not always correlate with superior outcomes. The DLinear model, a relatively simple architecture, delivered competitive long-term results with fewer computational demands.

Furthermore, the findings emphasize the importance of evaluating both technical and financial performance. Models that are accurate in prediction do not always lead to profitable trading outcomes, and vice versa. Thus, comprehensive, multi-dimensional evaluation frameworks, like the one introduced in this study, are critical for practical deployment.

Integrating dynamic or adaptive trading strategies, exploring ensemble methods, and incorporating rolling-window validation could further enhance model reliability. The study also proposes broader use of non-parametric statistical methods to validate findings across diversified market conditions.