Data-Efficient Realized Volatility Forecasting with Vision Transformers

AI summary

Overview

Research area: Financial machine learning — deep learning applied to options data, specifically forecasting volatility from implied volatility (IV) surfaces.

Technical level: Intermediate. Readers need some familiarity with options/volatility concepts (implied volatility surfaces, delta, days-to-maturity, realized volatility) and with transformer/Vision Transformer terminology, but the paper is a short, preliminary study rather than a heavily theoretical work.

Scope in one sentence: The paper is a preliminary study that trains Vision Transformer models (46,466 to 1,732,610 parameters) to predict an asset's realized volatility over the next 30 days from a single day's implied volatility surface, augmented with date information, using 4,259,070 rows of OptionMetrics data covering 2012–2022.

What This Paper Is About

Machine learning in finance often struggles because market data is noisy and scarce — the authors note that their entire preprocessed dataset is only 6.1 GB, while the text corpora used to train frontier LLMs contain multiple terabytes. At the same time, prior work has argued for the "virtue of complexity": deep models can capture highly nonlinear relationships that simpler factor models miss.

The goal is to test whether the Vision Transformer (ViT) architecture — normally used for image recognition — can be repurposed to read a day's implied volatility surface (treated as a single-channel image) and predict the realized volatility of the underlying asset over the following 30 days. The authors explicitly frame this as a first step toward a transformer model for options data, a setting they say remains largely unexplored.

Key Contributions

1. A demonstrated application of Vision Transformers to IV surfaces. The authors treat the smoothed, interpolated OptionMetrics IV surface as a small single-channel image of size 10×36 and adapt the standard PyTorch ViT to accept tensors of shape ℝ^(1×10×36) rather than square RGB images, with 2×2 patches.

2. A data-efficiency study across model sizes and training-set lengths. They compare six ViT configurations (ViT_0.005M_wide, ViT_0.12M_deep, ViT_0.17M_wide, ViT_0.5M_deep, ViT_0.5M_wide, ViT_1.7M) trained on one, four, or ten years of data, mapping out how model size interacts with available data.

3. Two ablation experiments. One removes the seasonality (month/day/day-of-week) augmentation; the other replaces the ViT with an MLP-only model of roughly matched parameter count.

4. A documented training recipe adapted from foundation-model practice. The procedure borrows components used to train text models such as DeepSeek-v3 and Llama — cosine annealing with warmup, AdamW, batch size 2048, Huber loss, Xavier initialization, batch normalization — scaled down to the available financial data.

Main Findings

• Best result: The best model is ViT_1.7M trained on 2012–2021 data, achieving R² = 0.41 on the 2022 test set.

• Small models are surprisingly capable: Models in the 0.05M–0.17M range can perform well when trained on small datasets but "collapse" on large ones. The 0.5M models do not improve on the smaller models. The 1.7M model gives the best performance but requires the full ten years of training data.

• ViTs beat MLPs at matched scale, and scale hurts the MLP: In the MLP-only ablation trained on four years (2018–2021) and tested on 2022, ViT_0.17M_wide reached R² = 0.37 and ViT_0.5M_wide reached R² = 0.37, versus 0.29 for both MLP_0_17 and MLP_0_12 and 0.17 for MLP_0_5 — larger MLP-only models performed worse.

• The tiny MLP only narrowly edges the tiny ViT: ViT_0.12M_deep reached R² = 0.27 in that ablation while the matched MLP_0_12 reached 0.29; across the larger pairs the ViT is clearly ahead.

• Seasonality augmentation helps only slightly: Removing month/day/day-of-week information had a small negative effect — ViT_0.5M_deep unchanged at 0.35, ViT_0.5M_wide down from 0.37 to 0.35, and ViT_1.7M down from 0.41 to 0.38. The authors read this as evidence that the model is primarily extracting nonlinear patterns from the IV surface itself.

• Training is fast to converge: All ViT models reach their maximum R² within one or two epochs, with further training causing overfitting.

• Market regime shifts hurt every model: All models show reduced performance when tested on 2020 data, which the authors attribute to market disruption during the COVID pandemic. Performance also degrades when training and test data are dissimilar, and one-year training runs can produce good but inconsistent results.

• Small models recover faster after a shock: Because they can be retrained cheaply as new data arrives, the small models provide an advantage in unstable conditions; the authors note the small models recover their performance on the 2021 test sample.

• The MLP baseline was harder to train: It required early stopping, batch normalization, and multiple training attempts with the best model selected at the end.

Methodology in Plain English

The authors pull two things from the OptionMetrics IvyDB database: the smoothed, interpolated implied volatility surface (a grid indexed by option delta on one axis and days to maturity on the other) and OptionMetrics' own realized volatility computed over 28 calendar days using the standard deviation of daily log returns. They split the data by year and month, drop incomplete samples, and end up with 4,259,070 rows covering 2012–2022.

Rather than feeding the surface to the model as a 3D object, they flatten it conceptually into a small grayscale-style image and add three extra values — month, day, and day of the week, all scaled to between 0 and 1 — so the model can pick up seasonal patterns. A month-by-month mapping between OptionMetrics' secid and CRSP's cusip keys was constructed using WRDS link tables so that future return-prediction work is possible.

The model is a standard Vision Transformer that cuts the 10×36 input into 2×2 patches, followed by a four-layer multilayer perceptron that turns the ViT's vector output into a single number. The authors vary depth (one or four MLP layers in the transformer module, giving "wide" versus "deep" variants) and width (hidden dimensions of 64, 128, or 256). Test sets are built so that predicting year y_i means training on the preceding n years.

Training follows the pattern of downsized foundation-model training: a cosine annealing learning rate schedule with warmup (first-cycle steps 200, max learning rate 0.01, min 0.001, warmup steps 100, gamma 0.95), the AdamW optimizer, batch size 2048 for all experiments, and Huber loss to limit the influence of outliers. The best epoch is selected, and R² on unseen test data is reported.

Why This Matters

Impact on research. Most prior deep-learning work on options data uses hand-constructed features or CNNs; this paper shows that a ViT — an architecture with different computational and training-stability properties — can extract nonlinear structure from IV surfaces directly. It also provides a concrete data-efficiency curve relating model size to training-set length, which is a question the field has mostly discussed informally.

Real-world applications:

• Volatility forecasting for risk management, where a 30-day-ahead realized volatility estimate feeds margin, VaR, and capital calculations.
• Options market making and pricing, where a fast, learned read of the entire IV surface could complement or challenge parametric surface models.
• Systematic trading and portfolio construction, since better volatility forecasts affect position sizing and hedging decisions.
• Retraining pipelines, since the paper shows small models can be retrained on as little as one year of data — practical when market regimes shift and older data becomes less relevant.

Industry relevance. The paper's central practical claim is that architecture and training design should be fit to the data available rather than to a benchmark leaderboard. For a quant desk, a model that trains on one year of data and recovers after a shock may be more useful than a larger model that needs a decade of history and degrades when conditions change.

Future Directions

1. Transfer learning on IV surfaces. The authors call this their top priority: fine-tuning a pretrained model, or retraining only the final regressor/classifier layers, to predict a different target. They note ViTs trained on image datasets exhibit this property and are often fine-tuned for medical or scientific imaging tasks.

2. Generalizing the ViT's learned representation to other prediction tasks. They tested whether retraining only the MLP predictor could support a new target — predicting the asset's return over the next 28 days — and did not observe the capability. They note this task is harder than predicting realized volatility.

3. Ensembling versus Mixture-of-Experts. Following their foundation-model framing, they propose comparing the ensembling approach used by prior options work against a Mixture-of-Experts architecture.

4. Theory of information content in IV surfaces. Since the link between model size, dataset size, and performance is clearly present but not explained, the authors suggest that a theoretical understanding of how much information IV surfaces carry could guide optimal data sampling.

Target Audience

This paper is most useful to quantitative researchers and machine learning practitioners working on financial time series — particularly those on options desks or in volatility research — who want a concrete, reproducible starting point for applying transformer architectures to IV surfaces. It also suits academic readers interested in data-efficient deep learning under scarce, noisy data, and graduate students looking for a clearly scoped preliminary study with an explicit list of open questions. Readers without exposure to options terminology will need background reading first, since the paper assumes familiarity with delta, days-to-maturity, and realized volatility.

Authors’ abstract

Recent work in financial machine learning has shown the virtue of complexity: the phenomenon by which deep learning methods capable of learning highly nonlinear relationships outperform simpler approaches in financial forecasting. While transformer architectures like Informer have shown promise for financial time series forecasting, the application of transformer models for options data remains largely unexplored. We conduct preliminary studies towards the development of a transformer model for options data by training the Vision Transformer (ViT) architecture, typically used in modern image recognition and classification systems, to predict the realized volatility of an asset over the next 30 days from its implied volatility surface (augmented with date information) for a single day. We show that the ViT can learn seasonal patterns and nonlinear features from the IV surface, suggesting a promising direction for model development.