Optimal decision-making in social settings is often based on forecasts from time series (TS) data. Recently, several approaches using deep neural networks (DNNs) such as recurrent neural networks (RNNs) have been introduced for TS forecasting and have shown promising results. However, the applicability of these approaches is being questioned for TS settings where there is a lack of quality training data and where the TS to forecast exhibit complex behaviors. Examples of such settings include financial TS forecasting, where producing accurate and consistent long-term forecasts is notoriously difficult. In this work, we investigate whether DNN-based models can be used to forecast these TS conjointly by learning a joint representation of the series instead of computing the forecast from the raw time-series representations. To this end, we make use of the dynamic factor graph (DFG) to build a multivariate autoregressive model. We investigate a common limitation of RNNs that rely on the DFG framework and propose a novel variable-length attention-based mechanism (ACTM) to address it. With ACTM, it is possible to vary the autoregressive order of a TS model over time and model a larger set of probability distributions than with previous approaches. Using this mechanism, we propose a self-supervised DNN architecture for multivariate TS forecasting that learns and takes advantage of the relationships between them. We test our model on two datasets covering 19 years of investment fund activities. Our experimental results show that the proposed approach significantly outperforms typical DNN-based and statistical models at forecasting the 21-day price trajectory. We point out how improving forecasting accuracy and knowing which forecaster to use can improve the excess return of autonomous trading strategies.

中文翻译：

---

用于金融时间序列长期预测的时空自适应神经网络

社会环境中的最佳决策通常基于时间序列 (TS) 数据的预测。最近，已经引入了几种使用深度神经网络 (DNN) 的方法，例如循环神经网络 (RNN)，用于 TS 预测，并显示出有希望的结果。然而，这些方法在 TS 设置中的适用性受到质疑，其中缺乏高质量的训练数据，并且要预测的 TS 表现出复杂的行为。此类设置的示例包括金融 TS 预测，众所周知，在其中生成准确且一致的长期预测非常困难。在这项工作中，我们研究了基于 DNN 的模型是否可以通过学习序列的联合表示而不是从原始时间序列表示计算预测来联合预测这些 TS。为此，我们利用动态因子图（DFG）来构建多元自回归模型。我们调查了依赖于 DFG 框架的 RNN 的一个共同局限性，并提出了一种新的基于可变长度注意力的机制 (ACTM) 来解决它。使用 ACTM，可以随时间改变 TS 模型的自回归阶数，并对比以前的方法更大的概率分布集进行建模。使用这种机制，我们提出了一种用于多变量 TS 预测的自监督 DNN 架构，该架构可以学习并利用它们之间的关系。我们在涵盖 19 年投资基金活动的两个数据集上测试我们的模型。我们的实验结果表明，所提出的方法在预测 21 天价格轨迹方面明显优于典型的基于 DNN 的统计模型。