Abstract:Uncertainty quantification is crucial in time series prediction, and quantile regression offers a valuable mechanism for uncertainty quantification which is useful for extreme value forecasting. Although deep learning models have been prominent in multi-step ahead prediction, the development and evaluation of quantile deep learning models have been limited. We present a novel quantile regression deep learning framework for multi-step time series prediction. In this way, we elevate the capabilities of deep learning models by incorporating quantile regression, thus providing a more nuanced understanding of predictive values. We provide an implementation of prominent deep learning models for multi-step ahead time series prediction and evaluate their performance under high volatility and extreme conditions. We include multivariate and univariate modelling, strategies and provide a comparison with conventional deep learning models from the literature. Our models are tested on two cryptocurrencies: Bitcoin and Ethereum, using daily close-price data and selected benchmark time series datasets. The results show that integrating a quantile loss function with deep learning provides additional predictions for selected quantiles without a loss in the prediction accuracy when compared to the literature. Our quantile model has the ability to handle volatility more effectively and provides additional information for decision-making and uncertainty quantification through the use of quantiles when compared to conventional deep learning models.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: in time - series prediction, although existing deep - learning models perform well in multi - step prediction, they are insufficient in providing comprehensive uncertainty quantification. To bridge this gap, the author proposes a new deep - learning framework based on quantile regression to improve the prediction ability for extreme values and highly volatile data and provide more detailed prediction information for decision - making.
### Specific problem description:
1. **Uncertainty Quantification**:
- In time - series prediction, uncertainty quantification is crucial for making more informed decisions, especially in highly volatile fields such as financial markets, energy demand, and weather forecasting.
- Traditional deep - learning models are good at multi - step prediction but often fail to provide comprehensive uncertainty measures.
2. **Extreme Value Prediction**:
- Extreme value prediction focuses on predicting rare and important events, which are usually outliers or extreme values in the dataset. Although they have a low probability of occurrence, they may have significant consequences.
- Existing deep - learning models perform poorly in handling extreme values, especially under high - volatility and extreme conditions.
3. **Multi - step Prediction**:
- Multi - step - ahead time - series prediction refers to predicting data at multiple future time points, which is more complex than single - step prediction, especially when the data has non - linear and multi - variable characteristics.
- When dealing with such tasks, deep - learning models need to pay special attention to the selection of model architectures and training methods to avoid the accumulation of prediction errors.
### Solutions:
- **Combination of Quantile Regression and Deep Learning**:
- By introducing quantile regression, this framework can predict conditional quantiles, thus providing more detailed prediction information, covering prediction values at different quantiles.
- The quantile loss function is used to optimize the model, so that the model can predict not only the average value but also values at specific quantiles, such as the 5%, 25%, 50%, 75%, and 95% quantiles.
- **Model Selection and Evaluation**:
- Advanced deep - learning models such as bidirectional long - short - term memory networks (BD - LSTM), convolutional long - short - term memory networks (Conv - LSTM), and encoder - decoder long - short - term memory networks (ED - LSTM) are used for experiments.
- By comparing the performance on the benchmark dataset, the prediction performance of these models under high - volatility and extreme conditions is evaluated.
- **Application Examples**:
- This framework has been tested in cryptocurrency markets such as Bitcoin and Ethereum, using daily closing price data for multi - step prediction, verifying its effectiveness and advantages in practical applications.
Through the above methods, this paper aims to improve the ability of deep - learning models to handle uncertainty and extreme values in time - series prediction, thereby providing more reliable prediction tools for fields such as finance and meteorology.