Abstract:Deep Learning is a consolidated, state-of-the-art Machine Learning tool to fit a function when provided with large data sets of examples. However, in regression tasks, the straightforward application of Deep Learning models provides a point estimate of the target. In addition, the model does not take into account the uncertainty of a prediction. This represents a great limitation for tasks where communicating an erroneous prediction carries a risk. In this paper we tackle a real-world problem of forecasting impending financial expenses and incomings of customers, while displaying predictable monetary amounts on a mobile app. In this context, we investigate if we would obtain an advantage by applying Deep Learning models with a Heteroscedastic model of the variance of a network's output. Experimentally, we achieve a higher accuracy than non-trivial baselines. More importantly, we introduce a mechanism to discard low-confidence predictions, which means that they will not be visible to users. This should help enhance the user experience of our product.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is: in regression tasks, deep - learning models only provide point estimates of the target and do not consider the uncertainty of prediction. This is a significant limitation in tasks that need to convey prediction confidence, especially in the financial field, where wrong predictions may bring risks. Specifically, the author focuses on the problem of predicting customers' upcoming financial expenditures and incomes through mobile applications. Due to the short and noisy time - series data, traditional deep - learning models perform poorly in this situation. To solve this problem, the author proposes a deep - learning method combined with a heteroscedastic model to better capture the uncertainty in prediction. ### Summary of the core problems in the paper: 1. **Prediction uncertainty**: Traditional deep - learning models only provide point estimates in regression tasks and do not consider the uncertainty of prediction. 2. **Short and noisy time - series**: The time - series of financial expenditures and incomes are usually short (only a few years of data), and there is a large amount of noise in the data. 3. **Improving prediction accuracy**: The author hopes to improve the accuracy of prediction by introducing uncertainty modeling and be able to screen out high - confidence prediction results, thereby enhancing the user experience. ### Solutions: - **Heteroscedastic model**: By introducing the heteroscedastic model, the model not only outputs the predicted value but also the variance of the predicted value, thereby evaluating the uncertainty of prediction. - **Rejecting low - confidence predictions**: By evaluating the uncertainty of prediction, those predictions with low confidence can be rejected to avoid showing unreliable information to users. ### Experimental results: Experiments show that the deep - learning method using the heteroscedastic model is superior to other benchmark methods in prediction accuracy and can effectively select high - confidence prediction samples, thereby improving the overall performance of the system. ### Related formulas: - The loss function of the heteroscedastic model is defined as: \[ L(w, b_{\text{ale}}; \{(x_i, y_i)\}_{i = 1}^N)=-\sum_{i = 1}^N\left[-\log(b_{\text{ale}})-\frac{\vert y_i-\phi(x_i)\vert}{b_{\text{ale}}}\right] \] where \(\phi(x)\) is the prediction function of the neural network, and \(b_{\text{ale}}\) is the scale parameter of the Laplace distribution. Through this method, the author has successfully solved the problem of short and noisy time - series prediction and improved the reliability of prediction and the user experience.

Uncertainty Modelling in Deep Networks: Forecasting Short and Noisy Series

A survey on uncertainty quantification in deep learning for financial time series prediction

Accurate and Reliable Forecasting using Stochastic Differential Equations

Prediction Under Uncertainty with Error-Encoding Networks

Predicting Weather Uncertainty with Deep Convnets

Modeling the Uncertainty in Electronic Health Records: a Bayesian Deep Learning Approach

Quantifying Uncertainty in Deep Learning Classification with Noise in Discrete Inputs for Risk-Based Decision Making

Forecasting Workload in Cloud Computing: Towards Uncertainty-Aware Predictions and Transfer Learning

A Benchmark on Uncertainty Quantification for Deep Learning Prognostics

Exploring Uncertainty in Deep Learning for Construction of Prediction Intervals

Predicting weather forecast uncertainty with machine learning

SDE-HNN: Accurate and Well-calibrated Forecasting using Stochastic Differential Equations

Understanding Uncertainty in Bayesian Deep Learning

Accurate Uncertainties for Deep Learning Using Calibrated Regression

Prediction of Energy Consumption for Variable Customer Portfolios Including Aleatoric Uncertainty Estimation

Single-Model Uncertainties for Deep Learning

Customer Lifetime Value Prediction with Uncertainty Estimation Using Monte Carlo Dropout

Probabilistic Deep Learning to Quantify Uncertainty in Air Quality Forecasting

Model Free Prediction with Uncertainty Assessment

Confidence Intervals and Simultaneous Confidence Bands Based on Deep Learning

Bayesian Recurrent Neural Network Models for Forecasting and Quantifying Uncertainty in Spatial-Temporal Data