Paper: arXiv 2512.12499

Authors: Pablo Hidalgo, Julio E. Sandubete, Agustín García-García

Abstract

This study investigates the contribution of Intrinsic Mode Functions (IMFs) derived from economic time series to the predictive performance of neural network models, specifically Multilayer Perceptrons (MLP) and Long Short-Term Memory (LSTM) networks. To enhance interpretability, DeepSHAP is applied, which estimates the marginal contribution of each IMF while keeping the rest of the series intact. Results show that the last IMFs, representing long-term trends, are generally the most influential according to DeepSHAP, whereas high-frequency IMFs contribute less and may even introduce noise, as evidenced by improved metrics upon their removal. Differences between MLP and LSTM highlight the effect of model architecture on feature relevance distribution, with LSTM allocating importance more evenly across IMFs.

Complexity vs Empirical Score

  • Math Complexity: 6.5/10
  • Empirical Rigor: 4.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper employs advanced mathematical concepts including Empirical Mode Decomposition, neural network architectures (MLP/LSTM), and formal SHAP value derivations with iterative formulas. While it uses real economic time series data and standard libraries for implementation, the description focuses more on methodological framework and interpretability results rather than providing extensive backtesting details, performance metrics, or direct trading implications.

Research Flowchart

  flowchart TD
  A["Research Goal<br>Explain IMF contributions<br>to NN forecasting"] --> B["Preprocessing<br>Decompose economic series into IMFs"]
  B --> C["Modeling<br>Train MLP and LSTM models"]
  C --> D["Explainability<br>Apply DeepSHAP to quantify IMF importance"]
  D --> E["Analysis & Outcomes<br>Long-term trend IMFs dominate importance"]
  D --> F["Analysis & Outcomes<br>High-frequency IMFs add noise & may be removed"]
  D --> G["Analysis & Outcomes<br>LSTM shows more even importance distribution"]
  E & F & G --> H["Conclusion<br>IMFs provide interpretable feature basis<br>for NN time series forecasting"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes to assess how Intrinsic Mode Functions (IMFs) from economic time series affect the predictive performance of MLP and LSTM neural networks, using DeepSHAP for interpretability.

Source sentence
This study investigates the contribution of Intrinsic Mode Functions (IMFs) derived from economic time series to the predictive performance of neural network models, specifically Multilayer Perceptrons (MLP) and Long Short-Term Memory (LSTM) networks.

Assumes 2

  • DeepLIFT (and thus DeepSHAP) assumes independence among features and local linearity. model
    Source sentence
    DeepLIFT propagates relevance scores backward through the network using linear rules under the assumptions of independence among features and local linearity.
  • The baseline input for DeepLIFT is chosen as the expected value of the inputs. model
    Source sentence
    If the baseline 𝐱(is chosen as the expected value of the inputs 𝐄[𝐱], DeepLIFT can be interpreted as an approximation to SHAP values under assumptions of independence and local linearity.

Data

  • NVIDIA historical financial records market-prices unknown
    May 2004 to May 2024 · NVIDIA
    Source sentence
    Regarding the primary dataset considered corresponds to NVIDIA’s historical financial records. It contains 5,219 time-ordered observations spanning two decades, from May 2004 to May 2024, decomposed into 9 IMFs.
  • Apple stock price series market-prices unknown
    May 2004 to May 2024 · Apple
    Source sentence
    A second dataset used in the analysis corresponds to Apple’s stock price series, covering the same period—from May 2004 to 2024—with 5,218 data points and decomposed into 8 IMFs.

Costs

Not applicable The paper reports no trading or execution result.

Finds 4

  • The MLP model achieves very high forecasting accuracy on the NVIDIA series when all IMFs are used, with MSE 13.605, RMSE 3.688, MAE 1.553 and R² 0.999. predictive-accuracy
    Source sentence
    Table 1. NVIDIA results. NVIDIA MSE 13.605 RMSE 3.688 MAE 1.553 R2 0.999
  • DeepSHAP attribution shows IMF 9 contributes the most (70.4% importance), followed by IMF 8 (14.4%). statistical
    Source sentence
    Figure 4 shows that IMF 9 has the highest contribution, with 70.4%. At a considerable distance, IMF 8 follows with 14.4% importance.
  • Long‑term (last) IMFs are generally the most influential according to DeepSHAP, while high‑frequency (first) IMFs have limited impact and may introduce noise. theoretical
    Source sentence
    which represent the long-term trends of the signal, tend to be the most relevant according to DeepSHAP, as they capture the overall dynamics of the time series.
  • Compared to MLP, the LSTM distributes relevance more evenly across IMFs, showing less bias toward high‑trend components. performance
    Source sentence
    The LSTM model, due to its memory capabilities and ability to capture temporal dependencies, distributes relevance more evenly across IMFs, showing less bias toward the high-trend components and assigning some importance to intermediate IMFs.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 45 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (18)

  1. COVID-19 Media Chatter and Macroeconomic Reflectors on Black Swan: A Spanish and Indian Stock Markets Comparison (2023) — link
  2. Analysis of financial pressure impacts on the health care industry with an explainable machine learning method: China versus the USA (2022) — link
  3. TS-MULE: Local Interpretable Model-Agnostic Explanations for Time Series Forecast Models (2021) — link
  4. Understanding Global Feature Contributions With Additive Importance Measures (2020)
  5. Deep Learning for Time Series Forecasting: Tutorial and Literature Survey (2020) — link
  6. From Local Explanations to Global Understanding with Explainable AI for Trees (2020) — link
  7. AR-Net: A simple Auto-Regressive Neural Network for time-series (2019) — link
  8. Recurrent Neural Networks for Time Series Forecasting: Current Status and Future Directions (2019) — link
and 10 more
  1. Performance Metrics (Error Measures) in Machine Learning Regression, Forecasting and Prognostics: Properties and Typology (2018) — link
  2. A Survey of Methods for Explaining Black Box Models (2018) — link
  3. Methods for interpreting and understanding deep neural networks (2017) — link
  4. Axiomatic Attribution for Deep Networks (2017) — link
  5. Explainable AI for Time Series Classification: A Review, Taxonomy and Research Directions (2022) — link
  6. Investigating explainablity methods in recurrent neural network architectures for financial time-series data [Master’s Thesis, University of the Witwatersrand, Johannesburg (South Africa)] (2022)
  7. Explainable Artificial Intelligence: Interpreting Default Forecasting Models Based on Machine Learning (SSRN Scholarly Paper No. 4090707) (2022)
  8. Predicting the Future with Artificial Neural Network (2018) — link
  9. A survey on feature selection methods (2014) — link
  10. Association for Computational Linguistics " Why Should I Trust You? " Explaining the Predictions of Any Classifier

0 of 18 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.