Paper: arXiv 2402.14476
Authors: Steven Y. K. Wong, Jennifer S. K. Chan, Lamiae Azizi
Abstract
Time-series with volatility clustering pose a unique challenge to uncertainty quantification (UQ) for returns forecasts. Methods for UQ such as Deep Evidential regression offer a simple way of quantifying return forecast uncertainty without the costs of a full Bayesian treatment. However, the Normal-Inverse-Gamma (NIG) prior adopted by Deep Evidential regression is prone to miscalibration as the NIG prior is assigned to latent mean and variance parameters in a hierarchical structure. Moreover, it also overparameterizes the marginal data distribution. These limitations may affect the accurate delineation of epistemic (model) and aleatoric (data) uncertainties. We propose a Scale Mixture Distribution as a simpler alternative which can provide favorable complexity-accuracy trade-off and assign separate subnetworks to each model parameter. To illustrate the performance of our proposed method, we apply it to two sets of financial time-series exhibiting volatility clustering: cryptocurrencies and U.S. equities and test the performance in some ablation studies.
Complexity vs Empirical Score
- Math Complexity: 8.5/10
- Empirical Rigor: 7.0/10
- Quadrant: Holy Grail — high math complexity, high empirical rigor
Why this score: The paper introduces a novel Scale Mixture Distribution for uncertainty quantification, involving advanced hierarchical probability theory and the interplay between aleatoric and epistemic uncertainties (high math). It validates the method with concrete backtests on financial datasets (cryptocurrencies and U.S. equities) and performs ablation studies, demonstrating implementation-heavy empirical rigor.
Research Flowchart
flowchart TD
A["Research Goal:<br>Quantify Neural Network<br>Uncertainty under<br>Volatility Clustering"] --> B{"Identify Limitations<br>of Deep Evidential Regression"}
B --> C["Proposed Method:<br>Scale Mixture Distribution"]
C --> D["Data & Experiments:<br>Cryptos & US Equities<br>+ Ablation Studies"]
D --> E["Computational Process:<br>Separate Subnetworks for<br>Model Parameters"]
E --> F["Key Findings:<br>Favorable Complexity-Accuracy<br>Trade-off & Better UQ"]
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 a Scale Mixture Distribution (SMD) with a prior only on the scaling factor ν and uses separate subnetworks for each distribution parameter, combined with model averaging and log‑squared returns as features.
Source sentence
We propose to place a prior solely on the scaling factor ν using a SMD, rather than on both mean and variance using the NIG prior.
Assumes 4
- The latent scaling factor ν must be greater than 1 to ensure a finite mean of the marginal distribution. model
Source sentence
We require α > 1 to ensure the mean of the marginal distribution is finite.
- Volatility clustering is present in the time‑series, making predictive uncertainty tracking essential. market
Source sentence
Time-series with volatility clustering pose a unique challenge to uncertainty quantification (UQ) for returns forecasts.
- Log of squared returns are informative features for modeling time‑varying forecast uncertainty. data
Source sentence
We propose to include the log of squared returns {log(r2 t‑K+1), . . . , log(r2 t )} as part of the input.
- Separate subnetworks for each hyperparameter improve specialization and uncertainty estimation. model
Source sentence
We show that modelling hyperparameters of the posterior distribution using subnetworks rather than a single output layer (as per Ensemble and Evidential) results in superior performance.
Data
- Cryptocurrency dataset crypto-onchain public
Source sentence
Our cryptocurrency dataset consists of hourly returns downloaded from Binance over July 2018 to December 2021, for ten of the most liquid, non‑stablecoin
- U.S. equities dataset market-prices unknown
Source sentence
A comparison in U.S. equities is also provided which illustrates performance in conventional financial time-series.
Costs
Not applicable
Finds 4
- SMD achieves superior RMSE and NLL performance compared to NIG on the UCI benchmark, winning on 6 of 9 datasets for RMSE and 8 of 9 for NLL. performance
Source sentence
We observe that both RMSE and NLL are overwhelmingly in SMD’s favor, achieving superior results in 6 (of 9) and 8 (of 9) of the datasets for RMSE and NLL, respectively.
- Separate subnetworks for each hyperparameter are essential for predictive uncertainty that closely tracks realized forecast error. predictive-accuracy
Source sentence
... that separate modelling of the hyperparameters is essential in producing predictive uncertainty estimates that closely track actual realized forecast error of Bitcoin and Chevron.
- Model averaging (ensemble) significantly improves return forecast accuracy without substantially increasing predictive uncertainty. performance
Source sentence
we incorporate model averaging (in the spirit of ensemble) into our Combined method and show that it significantly improves return forecast accuracy without significantly increasing the estimated predictive uncertainty.
- The Combined method outperforms state‑of‑the‑art Evidential and Ensemble methods in return forecasting on both cryptocurrency and U.S. equities data. performance
Source sentence
superior return forecasting performance than the two state-of-the-art neural network UQ methods — Ev
Checked by others
No other paper in the archive has been found to support or contradict these findings (compared with 37 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 (95)
- End to End Learning for Self-Driving Cars (2016) — link
- Deep Learning (2016)
- Deep Evidential Regression (2020)
- Language Models are Few-Shot Learners (2020)
- Advances in Neural Information Processing Systems 33 (2020)
- Supervised Temporal Autoencoder for Stock Return Time-series Forecasting (2021)
- Proceedings of the IEEE 45th Annual Computer Software and Applications Conference ( COMPSAC ‘21) (2021)
- Bayesian theory (2000)
and 87 more
- The Exponential Family: Conjugate Priors (2009)
- Convex optimization (2004)
- Simple and Scalable Predictive Uncertainty Estimation Using Deep Ensembles (2017)
- What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? (2017)
- Advances in Neural Information Processing Systems 30 (2017)
- Bagging Predictors (1996)
- Online Prediction Under Model Uncertainty via Dynamic Model Averaging: Application to a Cold Rolling Mill (2010) — link
- Empirical properties of asset returns: stylized facts and statistical issues (2001)
- A practical guide to volatility forecasting through calm and storm (2011)
- Probabilistic Backpropagation for Scalable Learning of Bayesian Neural Networks (2015)
- Proceedings of the 32nd International Conference on Machine Learning (ICML) (2015)
- Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning (2016)
- Proceedings of the 33rd International Conference on Machine Learning (ICML) (2016)
- Cryptocurrencies change everything (2021) — link
- Investing with cryptocurrencies – evaluating their potential for portfolio allocation strategies (2021)
- Negative bubbles and shocks in cryptocurrency markets (2016)
- Testing for Bubbles in Cryptocurrencies with Time-Varying Volatility (2018) — link
- Sentiment-Induced Bubbles in the Cryptocurrency Market (2019)
- Statistical analysis of bitcoin during explosive behavior periods (2019) — link
- Profit Potential in an Almost Efficient Market (1974)
- Active Portfolio Management: A Quantitative Approach for Producing Superior Returns and Controlling Risk (1999)
- Pattern Recognition and Machine Learning (2006)
- A New Interpretation of Information Rate (1956)
- Portfolio Selection (1952)
- Asset Allocation: Combining Investor Views with Market Equilibrium (1991)
- A Survey of Uncertainty in Deep Neural Networks (2021)
- The Elements of Statistical Learning (2020)
- Neural Network–Based Financial Volatility Forecasting: A Systematic Review (2022)
- Bayesian Learning for Neural Networks (1996)
- Predictability of Stock Returns: Robustness and Economic Significance (1995)
- Long Short-Term Memory (1997)
- Autoregressive Conditional Heteroscedasticity with Estimates of the Variance of United Kingdom Inflation (1982)
- Generalized autoregressive conditional heteroskedasticity (1986) — link
- Time Series Analysis: Forecasting and Control (1994)
- A Comparison of ARIMA and LSTM in Forecasting Time Series (2018)
- Proceedings of the 17th IEEE International Conference on Machine Learning and Applications (2018)
- A Large-Scale Study of Probabilistic Calibration in Neural Network Regression (2023)
- Uncertainty Estimation by Fisher Information-based Evidential Deep Learning (2023)
- Proceedings of the 40th International Conference on Machine Learning (ICML) (2023)
- Novel volatility forecasting using deep learning–Long Short Term Memory Recurrent Neural Networks (2019)
- Generalized Framework For Applying The Kelly Criterion To Stock Markets (2018) — link
- Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods (2021)
- A Practical Bayesian Framework for Backpropagation Networks (1992) — link
- Uncertainty in Deep Learning (2016)
- Monte Carlo Sampling Methods Using Markov Chains and Their Applications (1970)
- On the Validity of Bayesian Neural Networks for Uncertainty Estimation (2019)
- Being Bayesian, Even Just a Bit, Fixes Overconfidence in ReLU Networks (2020)
- Can You Trust Your Model’s Uncertainty? Evaluating Predictive Uncertainty under Dataset Shift (2019)
- Advances in Neural Information Processing Systems 32 (2019)
- Speeding Up MCMC by Efficient Data Subsampling (2019) — link
- Scale mixtures of normal distributions (1974)
- Scale Mixtures Distributions in Statistical Modelling (2008)
- Empirical Asset Pricing via Machine Learning (2020) — link
- WaveNet: a deep convolutional neural network using directional wavelets for low-dose X-ray CT reconstruction (2017) — link
- Hands-On Bayesian Neural Networks — A Tutorial for Deep Learning Users (2022)
- The Unreasonable Effectiveness of Deep Evidential Regression (2022)
- Efficient and Robust LiDAR-Based End-to-End Navigation (2021)
- 2021 IEEE International Conference on Robotics and Automation (2021)
- Evidential Deep Learning for Guided Molecular Property Prediction and Discovery (2021)
- Vision-Based Autonomous Car Racing Using Deep Imitative Reinforcement Learning (2021) — link
- Leveraging Graph and Deep Learning Uncertainties to Detect Anomalous Maritime Trajectories (2022) — link
- 3D High-Quality Magnetic Resonance Image Restoration in Clinics Using Deep Learning (2022)
- On Second-Order Scoring Rules for Epistemic Uncertainty Quantification (2023)
- Sources of Uncertainty in Machine Learning – A Statisticians’ View (2023)
- Introduction to Uncertainty Quantification (2015)
- Aleatory and epistemic uncertainty in probability elicitation with an example from hazardous waste management: Treatment of aleatory and epistemic uncertainty (1996)
- The Well-Calibrated Bayesian (1982)
- The Comparison and Evaluation of Forecasters (1983)
- Lessons Learned in the Challenge: Making Predictions and Scoring Them (2006)
- Improving Evidential Deep Learning via Multi-Task Learning (2022)
- Natural Posterior Network: Deep Bayesian Predictive Uncertainty for Exponential Family Distributions (2021)
- 9th International Conference on Learning Representations (ICLR) (2021)
- Regression Prior Networks (2020)
- Strictly Proper Scoring Rules, Prediction, and Estimation (2007)
- GARCH Modeling of Stock Market Volatility
- Stock market volatility (2009)
- Adam: A Method for Stochastic Optimization (2015)
- 3rd International Conference on Learning Representations (2015)
- Generalized autoregressive conditional heteroskedasticity (1986)
- Mixture Density Networks (1994) — link
- Multivariate Deep Evidential Regression (2022) — link
- Uncertainty Regularized Evidential Regression (2024) — link
- Deep Evidential Learning for Bayesian Quantile Regression (2023) — link
- A Survey on Uncertainty Quantification Methods for Deep Learning (2024) — link
- A survey on uncertainty quantification in deep learning for financial time series prediction (2024) — link
- Analyzing return asymmetry and quantiles through stochastic volatility models using asymmetric Laplace error via uniform scale mixtures (2015)
- A new look at cryptocurrencies (2018)
0 of 95 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.