Paper: arXiv 2603.16886

Authors: Nabeel Ahmad Saidd

Abstract

Multi-horizon price forecasting is central to portfolio allocation, risk management, and algorithmic trading, yet deep learning architectures have proliferated faster than rigorous financial benchmarks can evaluate them. This study provides a controlled comparison of nine architectures (Autoformer, DLinear, iTransformer, LSTM, ModernTCN, N-HiTS, PatchTST, TimesNet, and TimeXer) spanning Transformer, MLP, CNN, and RNN families across cryptocurrency, forex, and equity index markets at 4-hour and 24-hour horizons. A total of 918 experiments were conducted under a strict five-stage protocol including fixed-seed Bayesian hyperparameter optimization, configuration freezing per asset class, multi-seed retraining, uncertainty aggregation, and statistical validation. ModernTCN achieves the best mean rank (1.333) with a 75 percent first-place rate, followed by PatchTST (2.000). Results reveal a clear three-tier ranking structure and show that architecture explains nearly all performance variance, while seed randomness is negligible. Rankings remain stable across horizons despite 2 to 2.5 times error amplification. Directional accuracy remains near 50 percent across all configurations, indicating that MSE-trained models lack directional skill at hourly resolution. The findings highlight the importance of architectural inductive bias over raw parameter count and provide reproducible guidance for multi-step financial forecasting.

Complexity vs Empirical Score

  • Math Complexity: 0.0/10
  • Empirical Rigor: 0.0/10
  • Quadrant: Philosophers — conceptual discussion, limited math and data

Why this score: Scoring failed: ’list’ object has no attribute ‘get’

Research Flowchart

  flowchart TD
    A["<b>Research Goal</b><br>Evaluate 9 Deep Learning Architectures for Multi-Horizon Financial Forecasting"] --> B["<b>Inputs & Assets</b><br>Crypto, Forex & Equity Indices<br><i>(4h & 24h Horizons)</i>"]
    
    B --> C["<b>5-Stage Methodology</b><br>Bayesian HPO, Config Freezing, Multi-Seed Retraining & Statistical Validation"]
    
    C --> D["<b>Execution</b><br>918 Controlled Experiments across Transformers, MLPs, CNNs & RNNs"]
    
    D --> E["<b>Key Findings: Architecture Tiering</b><br>1. ModernTCN (Rank 1.33, 75% Win Rate)<br>2. PatchTST (Rank 2.00)<br>3. Tiered performance driven by inductive bias"]
    
    D --> F["<b>Key Findings: Market Dynamics</b><br>• Variance driven by Architecture, not Seed<br>• Errors amplify 2-2.5x at 24h horizon<br>• Directional Accuracy remains ~50% (MSE loss)"]

Cited References (37)

  1. Time Series Analysis: Forecasting and Control (1970)
  2. Generalized Autoregressive Conditional Heteroskedasticity (1986)
  3. Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation (1982)
  4. Automatic Time Series Forecasting: The Forecast Package for R (2008)
  5. Long Short-Term Memory (1997)
  6. Learning to Forget: Continual Prediction with LSTM (2000)
  7. Attention is All You Need (2017)
  8. Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting (2021)
and 29 more
  1. A Time Series is Worth 64 Words: Long-term Forecasting with Transformers (2023)
  2. iTransformer : Inverted Transformers Are Effective for Time Series Forecasting (2024)
  3. TimeXer : Empowering Transformers for Time Series Forecasting with Exogenous Variables (2024)
  4. Are Transformers Effective for Time Series Forecasting? (2023)
  5. N-HiTS : Neural Hierarchical Interpolation for Time Series Forecasting (2023)
  6. TimesNet : Temporal 2D-Variation Modeling for General Time Series Analysis (2023)
  7. ModernTCN : A Modern Pure Convolution Structure for General Time Series Analysis (2024)
  8. The M4 Competition: Results, Findings, Conclusion and Way Forward (2018)
  9. The M5 Accuracy Competition: Results, Findings, and Conclusions (2022)
  10. Recurrent Neural Networks for Time Series Forecasting: Current Status and Future Directions (2021)
  11. Time-Series Forecasting with Deep Learning: A Survey (2021)
  12. A Review and Comparison of Strategies for Multi-step Ahead Time Series Forecasting Based on the NN5 Forecasting Competition (2012)
  13. Direct Multi-Step Estimation and Forecasting (2007)
  14. The Use of Ranks to Avoid the Assumption of Normality Implicit in the Analysis of Variance (1937)
  15. Individual Comparisons by Ranking Methods (1945)
  16. Statistical Comparisons of Classifiers over Multiple Data Sets (2006)
  17. Distribution-Free Multiple Comparisons (1963)
  18. Comparing Predictive Accuracy (1995)
  19. Testing the Equality of Prediction Mean Squared Errors (1997)
  20. Optuna: A Next-Generation Hyperparameter Optimization Framework (2019)
  21. Algorithms for Hyper-Parameter Optimization (2011)
  22. Financial Time Series Forecasting with Deep Learning: A Systematic Literature Review: 2005–2019 (2020)
  23. An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling (2018) — link
  24. Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift (2022)
  25. Deep Reinforcement Learning That Matters (2018)
  26. Accounting for Variance in Machine Learning Benchmarks (2021)
  27. PyTorch : An Imperative Style, High-Performance Deep Learning Library (2019)
  28. Empirical Properties of Asset Returns: Stylized Facts and Statistical Issues (2001)
  29. Efficient Capital Markets: A Review of Theory and Empirical Work (1970)

0 of 37 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2518 linked papers.