Paper: arXiv 2306.12446
Authors: Wenbo Ge, Pooia Lalbakhsh, Leigh Isai, Artem Lensky, Hanna Suominen
Abstract
This study aims to compare multiple deep learning-based forecasters for the task of predicting volatility using multivariate data. The paper evaluates a range of models, starting from simpler and shallower ones and progressing to deeper and more complex architectures. Additionally, the performance of these models is compared against naive predictions and variations of classical GARCH models. The prediction of volatility for five assets, namely S&P500, NASDAQ100, gold, silver, and oil, is specifically addressed using GARCH models, Multi-Layer Perceptrons, Recurrent Neural Networks, Temporal Convolutional Networks, and the Temporal Fusion Transformer. In the majority of cases, the Temporal Fusion Transformer, followed by variants of the Temporal Convolutional Network, outperformed classical approaches and shallow networks. These experiments were repeated, and the differences observed between the competing models were found to be statistically significant, thus providing strong encouragement for their practical application.
Complexity vs Empirical Score
- Math Complexity: 4.0/10
- Empirical Rigor: 7.0/10
- Quadrant: Street Traders — empirical and applied, lower math complexity
Why this score: The paper applies standard deep learning architectures (RNNs, CNNs, Transformers) and classical GARCH models with relatively low mathematical derivation, focusing on implementation and benchmarking. It demonstrates strong empirical rigor through backtesting on multiple assets, statistical significance testing, and comparison against baselines.
Research Flowchart
flowchart TD
A["Research Goal<br>Compare DL Models for<br>Multivariate Volatility Prediction"] --> B["Data Preparation<br>5 Assets: S&P500, NASDAQ100, Gold, Silver, Oil"]
B --> C["Model Implementation"]
subgraph C ["Computational Processes"]
C1["GARCH Models<br>Classical Baseline"]
C2["MLP<br>Shallow Network"]
C3["RNN<br>Recurrent Neural Net"]
C4["TCN<br>Temporal Convolutional Net"]
C5["TFT<br>Temporal Fusion Transformer"]
end
C --> D["Statistical Evaluation<br>Repeated Experiments & Significance Testing"]
D --> E{"Key Findings/Outcomes"}
E --> F["TFT & TCN Variants<br>Outperform All Others"]
E --> G["Significant Improvement<br>Over GARCH & Shallow Networks"]
E --> H["Strong Case for<br>Practical Application"]
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 compare multiple deep learning-based forecasters for volatility prediction using multivariate data, evaluating models from shallow MLPs to deep TFT and TCN architectures.
Source sentence
This study aims to compare multiple deep learning-based forecasters for the task of predicting volatility using multivariate data.
Assumes 6
- The current trading day is assumed to be over and observed when forecasting volatility. data
Source sentence
This assumed that at the time of forecasting, the current trading day is over and observed, an assumption maintained for all forecasting models.
- A 70-15-15 train‑validation‑test split is used for model evaluation. data
Source sentence
…ction of a 70-15-15 train-validation-test split with the evaluation process were statistically significantly different from each other.
- Statistical significance testing assumes normality of MSE distributions when using the Shapiro‑Wilk test. statistical
Source sentence
The Shapiro-Wilk (SW) test [58] was first applied to assess the normality of the distribution with significance level 𝛼= 0.05.
- If both distributions are normal, Student’s t‑test is used; otherwise Kruskal‑Wallis test is applied. statistical
Source sentence
If both distributions were normal then Student’s t-test [61] was used, otherwise the Kruskal-Wallis (KW) test [31] was employed.
Show 2 more
- The naive benchmark repeats the most recent known volatility value. model
Source sentence
The naïve model simply repeated the most recent known value of volatility (Fig. 2).
- GARCH variants are selected based on auto‑correlation and partial auto‑correlation functions. model
Source sentence
Lag orders were 1 and 2, as informed by the auto-correlation and partial auto-correlation functions.
Data
- Five asset price series market-prices unknown
Source sentence
The prediction of volatility for five assets, namely S&P500, NASDAQ100, gold, silver, and oil, is specifically addressed using GARCH models, Multi-Layer Perceptrons, Recurrent Neural Networks, Temporal Convolutional Networks, and the Temporal Fusion Transformer.
Costs
Not applicable
Finds 2
- The Temporal Fusion Transformer (TFT) most often achieved the lowest MSE, followed by variants of the Temporal Convolutional Network (TCN), outperforming classical GARCH and shallow networks with statistical significance. performance
Source sentence
In the majority of cases, the Temporal Fusion Transformer, followed by variants of the Temporal Convolutional Network, outperformed classical approaches and shallow networks.
- Variants of TCN and TFT forecast volatility with less error across almost all assets and tasks investigated, with statistical significance. statistical
Source sentence
The results indicate that variants of TCN and TFT were able to forecast volatility with less error across almost all assets and tasks investigated, with statistical significance.
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 39 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.