Abstract

Compares ARMA models, boosting, neural network models, HAR_RV models and proposes a new method for predicting one day ahead realized volatility of financial series. HAR_RV models are taken as compared classical volatility prediction models. In addition, the phenomenon of transfer learning for boosting and neural network models is investigated. Bitcoin and E-mini S&P500 are chosen as examples. The realized volatility is calculated based on intraday (intraday—24 hours) data. The calculation is based on the closing values of the internal five-minute intervals. Comparisons are made both within and between the two intervals. The intervals considered are January 1, 2018–January 1, 2022 and January 1, 2018–April 2, 2023. Since there were structural changes in the markets during these intervals, the models are estimated in sliding windows of 399 days length. For each time series, we compare three-parameter enumeration boosting, about 10 different neural network architectures, ARMA models, the newly proposed CTCM method, and various training transfer and training sample expansion options. It is shown that ARMA and HAR_RV models are generally inferior to other listed methods and models. The CTCM model and neural networks of CNN architecture are the most suitable for financial time series forecasting and show the best results. Although transfer learning shows no improvement in terms of forecast precision and yields little decline. It requires more extensive and detailed study. The smallest MAPEs for Bitcoin and E-mini S&P500 realized volatility forecasts are achieved by the newly proposed CTCM model and are 21.075%, 25.311% on the first interval and 21.996%, 26.549% on the second interval, respectively.

中文翻译：

---

用于预测比特币和 E-Mini S&P500 实际波动性的修正三重校正方法、CNN 和迁移学习

 抽象的

比较了 ARMA 模型、boosting、神经网络模型、HAR_RV 模型，并提出了一种预测金融序列未来一天已实现波动性的新方法。 HAR_RV 模型被视为经典波动率预测模型的比较。此外，还研究了 boosting 和神经网络模型的迁移学习现象。选择比特币和E-mini S&P500作为例子。已实现波动率是根据日内（日内—24小时）数据计算的。该计算基于内部五分钟间隔的收盘值。在两个区间内和之间进行比较。考虑的时间间隔为2018年1月1日至2022年1月1日和2018年1月1日至2023年4月2日。由于这些时间间隔内市场发生结构性变化，因此模型在长度为399天的滑动窗口中进行估计。对于每个时间序列，我们比较了三参数枚举提升、大约 10 种不同的神经网络架构、ARMA 模型、新提出的 CTCM 方法以及各种训练转移和训练样本扩展选项。结果表明，ARMA 和 HAR_RV 模型通常不如其他列出的方法和模型。 CTCM模型和CNN架构的神经网络最适合金融时间序列预测，并且表现出最好的结果。尽管迁移学习在预测精度方面没有任何改进，而且产量也几乎没有下降。需要更广泛、更细致的研究。比特币和 E-mini S&P500 已实现波动率预测的最小 MAPE 是通过新提出的 CTCM 模型实现的，第一个区间的 MAPE 分别为 21.075%、25.311%，第二个区间的 MAPE 分别为 21.996%、26.549%。