Perturbation-Specific Reliability of Intrinsic and Post-Hoc Explainable Machine Learning in Volatility Forecasting
This study demonstrates that the reliability of both intrinsic and post-hoc explanations in volatility forecasting is highly perturbation-specific, varying significantly across different model–representation–explainer configurations and necessitating empirical evaluation under diverse conditions such as retraining, temporal shifts, and input noise.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the world of finance, predicting how much a stock price will swing is often considered easier than guessing which way the price will move. This measure of movement, known as volatility, is crucial for managing risk, yet it remains notoriously difficult to forecast because markets are messy, unpredictable, and constantly changing. To navigate this uncertainty, financial experts increasingly rely on machine learning, where computer programs learn from past data to make future predictions. However, these powerful programs often operate as "black boxes," meaning their internal logic is hidden and difficult for humans to understand. This lack of transparency creates a problem: if a computer makes a risky prediction, regulators and risk managers need to know why. To solve this, scientists have developed two main ways to make these machines explain themselves. The first approach builds the explanation directly into the machine's design from the start, creating models that are naturally easy to read. The second approach trains a complex, opaque machine first and then uses a separate tool to look at the finished product and guess what it was thinking. The big question has always been which of these two methods is actually more reliable.
A team of researchers from Universitas Ahmad Dahlan in Indonesia set out to answer this question by putting these two approaches to a rigorous test in the high-stakes environment of stock market forecasting. They did not simply ask which method looked better on paper; instead, they subjected the explanations to a series of real-world stress tests to see how well they held up when conditions changed. The researchers focused on predicting the five-day volatility of twenty major stocks from the United States, Europe, and Japan, using data spanning from 2016 to 2025. They trained five different types of machine learning models, pairing each with the explanation method best suited for its architecture. Crucially, they only included models in their final analysis if they proved they could actually predict volatility better than a simple, baseline guess based on recent history. This ensured they were only evaluating explanations for models that actually possessed genuine forecasting skill.
The study introduced a clever twist to isolate the source of reliability. For one specific type of machine learning model, the researchers applied both the "built-in" explanation and the "post-training" explanation to the exact same trained network. This allowed them to see if the difference in reliability came from the model itself or from the method used to explain it. They then tested these explanations under three different types of pressure. First, they added a small amount of random noise to the input data to see if the explanation remained steady when the numbers were slightly fuzzy. Second, they shifted the time window to see if the explanation held up as market conditions changed over months and years. Third, they retrained the model from scratch using different random starting points to see if the explanation remained consistent even when the training process varied slightly.
The results revealed a surprising truth: there is no single "best" way to explain a machine learning model. The reliability of an explanation depends entirely on the specific combination of the model, the data it sees, and the type of pressure it faces. In one of the most striking findings, the researchers discovered that for the same machine learning network, the built-in explanation was far more stable when the input data was noisy, but it fell apart completely when the model was retrained. Conversely, the explanation generated after training was less stable against noisy data but remained much more consistent when the model was rebuilt from scratch. This means that the choice of explanation method should not be based on a general rule about whether it is "built-in" or "added later," but rather on the specific task at hand. If a user needs an explanation that can withstand small errors in the data, the built-in method is superior. If they need an explanation that remains consistent even when the model is updated or retrained, the post-training method is the better choice.
The study also challenged a common assumption in the field regarding how we measure the quality of these explanations. Researchers often try to verify if an explanation is truthful by removing the most important features one by one to see if the model's prediction gets worse. The team found that in the context of stock volatility, this method failed to distinguish between a smart explanation and a random guess. Because the signal in financial data is so weak and the noise so high, removing a feature did not reliably change the outcome, making this standard test useless for this specific problem. Furthermore, the researchers found that a simple, traditional statistical model known as the Heterogeneous Autoregressive model of Realized Volatility (HAR-RV) achieved the best median forecasting accuracy in the study, decisively outperforming all the machine learning models tested. Remarkably, a spline-based Generalized Additive Model (GAM), which is intrinsically interpretable, performed just as well as this top-performing HAR-RV benchmark, while the more complex neural networks and tree-based models lagged behind. This suggests that for this specific task, the added complexity of deep learning did not buy any extra accuracy, and the simpler, more transparent models were just as effective.
Ultimately, the research concludes that we cannot assume an explanation is trustworthy just because it comes from a model designed to be transparent. The "built-in" explanations, which are often marketed as the gold standard for interpretability, proved to be the least reproducible when the model was retrained. Specifically, the intrinsic attention mask of the TabNet model showed near-zero seed reproducibility (with a rank agreement of only 0.006), meaning it changed its story entirely with a new set of random starting numbers, making it the least reliable explanation in the study. The study demonstrates that reliability is not an inherent property of a specific type of explanation, but a fragile characteristic that must be tested empirically for every new situation. For anyone relying on these tools to make financial decisions, the lesson is clear: you must verify that the explanation holds up under the specific pressures your system will face, rather than trusting it simply because of how the machine was built.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.