The researchers set out to answer a deceptively simple question: under controlled and fair conditions, how much of volatility forecasting performance is actually determined by model design, and how much by the inherent structure of financial time series? To find out, they built a rigorous experimental framework spanning 14 global equity indices, forecast horizons ranging from a single trading day all the way out to 100 trading days, and a diverse roster of competing models. The lineup included classical econometric workhorses such as GARCH and the Heterogeneous Autoregressive (HAR) model, alongside modern machine learning approaches including tree-based ensembles and deep neural networks equipped with attention mechanisms.

Crucially, the authors enforced what they call a capacity-controlled and strictly chronological evaluation protocol. All models were trained, validated, and tested on temporally ordered data with no leakage from future to past, and model complexity, measured in the number of free parameters, was deliberately varied across experiments. This design allowed the team to isolate the effect of capacity, the expressive power of a model, from other confounds that typically muddy comparisons between econometric and machine learning methods.

The results paint a striking picture of what financial volatility data actually looks like from a statistical standpoint. Post-hoc diagnostics revealed that realized volatility is organized around three dominant features: persistent regimes in which calm and turbulent periods stretch across long stretches of time, strong synchronization across markets, and sudden discontinuous shocks that arrive without warning. Global equity indices, in other words, do not behave like collections of independent observations. They are deeply and persistently dependent, with the past encoding far more about the future than a naive count of data points would suggest.

This structure has profound consequences for what machine learning models can learn. The study found that forecasting performance across all architectures was governed by a single principle: parameter identifiability. When the number of parameters in a model remained low relative to the effective information available in the training data, predictions stayed stable and reliable across forecast horizons. But once parameter counts grew beyond the information the data could support, even seemingly abundant datasets of thousands of daily observations, performance deteriorated sharply. The models did not fail because they lacked sophistication; they failed because there simply was not enough independent structure in the data to pin down so many free parameters.

Perhaps the most eye-catching finding concerns attention mechanisms, the innovations behind transformers that have revolutionized natural language processing and, increasingly, time series forecasting. In this study, attention did not discover any hidden structure in the volatility data that simpler models missed. Instead, the attention weights converged on what classical econometricians would recognize immediately: recent observations dominate. The neural networks, in effect, rediscovered the well-known persistence of volatility and settled into configurations that mimic the weighting schemes of much simpler models. The extra machinery, it turns out, was largely redundant.

The researchers’ central theoretical contribution is the concept of effective information as distinct from nominal sample size. For strongly dependent time series such as daily volatility, consecutive observations overlap heavily in the information they carry. A dataset of ten thousand daily observations does not provide ten thousand independent pieces of evidence; the long memory and regime persistence of volatility mean that the effective sample size is dramatically smaller. The study argues that model capacity must therefore be calibrated to this effective information content rather than to the raw number of observations, a reframing with direct implications for anyone applying deep learning to financial, climate, or other strongly autocorrelated data.

The methodological lineage of the study reaches back decades. The GARCH framework, introduced in the 1980s by Robert Engle and Tim Bollerslev, remains the backbone of conditional volatility modeling, capturing the tendency of large shocks to be followed by large shocks. The HAR model, proposed by Corsi in 2009, offers a parsimonious approximation of long memory by blending daily, weekly, and monthly volatility components. The realized volatility literature pioneered by Andersen, Bollerslev, Diebold, and Labys established that high-frequency data can yield far more precise measures of latent volatility. Against this classical foundation, the new study positions modern tools including XGBoost-style tree boosting, long short-term memory (LSTM) networks, and transformer-based architectures, some of which have previously reported impressive results in financial prediction tasks.

What distinguishes this work from prior benchmarking efforts is its insistence on fairness and interpretability of results. Many previous comparisons between econometric and machine learning models have been criticized for unequal tuning budgets, inconsistent data splits, or evaluation windows that favor one family of methods. By imposing strict chronology and systematically varying capacity, the KIIT team was able to show that the apparent superiority of complex neural models in earlier studies often evaporates under controlled conditions. Low-capacity models, particularly the HAR family, remained competitive across virtually all horizons, while high-capacity neural models added little beyond horizon-specific noise.

The implications extend well beyond academic benchmarking. For practitioners in risk management, portfolio construction, and derivatives pricing, the message is that a well-specified parsimonious model may outperform a heavyweight neural network, not because deep learning is inherently unsuited to finance, but because the data cannot support the complexity. For researchers designing neural forecasting systems, the study suggests a shift in focus: rather than adding parameters and layers, attention should turn to expanding the effective information available to models, whether through richer data sources, cross-market signals, or novel representations that break the dependency structure that limits learnability.

The study also offers a broader lesson about the scientific method in the era of deep learning. As machine learning spreads into domains governed by persistent, dependent processes, from finance to epidemiology to hydrology, the temptation to deploy large architectures grows. This research demonstrates that nominal sample size is a misleading measure of learnability in such settings, and that disciplined capacity control is not a constraint on progress but a prerequisite for it. The authors’ supplementary repository, hosted publicly on GitHub, includes all figures, tables, and diagnostic outputs, and all underlying market data were drawn from publicly accessible Yahoo Finance price series, making the entire evaluation pipeline reproducible by any interested reader.

The paper was received in April 2025 and published online on 29 August 2026 as part of Volume 38 of Neural Computing and Applications. The authors disclose no competing interests and note that AI-assisted tools were used only for editorial refinement of the manuscript, not for research design or analysis. In a field crowded with claims that the latest architecture finally cracks the market, this study stands out for what it does not claim: no hidden structure discovered, no breakthrough accuracy, just a carefully controlled demonstration that in the world of volatility forecasting, the data, not the model, sets the ceiling. Recognizing that ceiling, the authors argue, may be the most valuable forecasting insight of all.