Adaptive Scheduling of Correlated Tasks for Stock Return Forecasting
Article Qlib
Summary
This paper description presents a learnable scheduler for sequence-learning problems with related prediction tasks, such as forecasting returns at different future horizons. During training, the scheduler chooses an auxiliary task based on the current model state and minibatch. The model learns from the selected task, while the scheduler is updated to improve the model’s validation performance; the two components are trained through bi-level optimization.
The described experiments use historical transaction data for 300 CSI 300 stocks, with time-ordered training, validation, and test splits. Two experimental settings differ in their split dates, target-return definitions, and choice of auxiliary task sets. The document points to a paper for results under one setting and a Qlib benchmark page for the other, but gives no numerical results itself. Results may depend on the data version and preprocessing, which the document says differ between the original experiments and the current Qlib baseline. It describes a forecasting method, not evidence of deployable trading profitability.
Key ideas
- The scheduler selects among temporally related auxiliary prediction tasks during model training.
- Task selection uses information about the current model state and training minibatch.
- Bi-level optimization trains the scheduler to improve validation performance and the forecasting model to reduce training loss.
- The stock experiments use time-based splits and vary target definitions and auxiliary task sets across settings.
- The document provides no numerical performance results and notes differences in data and preprocessing between settings.
Tags
Full text
# Temporally Correlated Task Scheduling for Sequence Learning
# Temporally Correlated Task Scheduling for Sequence Learning
### Background
Sequence learning has attracted much research attention from the machine learning community in recent years. In many applications, a sequence learning task is usually associated with multiple temporally correlated auxiliary tasks, which are different in terms of how much input information to use or which future step to predict. In stock trend forecasting, as demonstrated in Figure1, one can predict the price of a stock in different future days (e.g., tomorrow, the day after tomorrow). In this paper, we propose a framework to make use of those temporally correlated tasks to help each other.
### Method
Given that there are usually multiple temporally correlated tasks, the key challenge lies in which tasks to use and when to use them in the training process. This work introduces a learnable task scheduler for sequence learning, which adaptively selects temporally correlated tasks during the training process. The scheduler accesses the model status and the current training data (e.g., in the current minibatch) and selects the best auxiliary task to help the training of the main task. The scheduler and the model for the main task are jointly trained through bi-level optimization: the scheduler is trained to maximize the validation performance of the model, and the model is trained to minimize the training loss guided by the scheduler. The process is demonstrated in Figure2.
<p align="center">
<img src="workflow.png"/>
</p>
At step <img src="https://latex.codecogs.com/png.latex?s" title="s" />, with training data <img src="https://latex.codecogs.com/png.latex?x_s,y_s" title="x_s,y_s" />, the scheduler <img src="https://latex.codecogs.com/png.latex?\varphi" title="\varphi" /> chooses a suitable task <img src="https://latex.codecogs.com/png.latex?T_{i_s}" title="T_{i_s}" /> (green solid lines) to update the model <img src="https://latex.codecogs.com/png.latex?f" title="f" /> (blue solid lines). After <img src="https://latex.codecogs.com/png.latex?S" title="S" /> steps, we evaluate the model <img src="https://latex.codecogs.com/png.latex?f" title="f" /> on the validation set and update the scheduler <img src="https://latex.codecogs.com/png.latex?\varphi" title="\varphi" /> (green dashed lines).
### Experiments
Due to different data versions and different Qlib versions, the original data and data preprocessing methods of the experimental settings in the paper are different from those experimental settings in the existing Qlib version. Therefore, we provide two versions of the code according to the two kinds of settings, 1) the [code](https://github.com/lwwang1995/tcts) that can be used to reproduce the experimental results and 2) the [code](https://github.com/microsoft/qlib/blob/main/qlib/contrib/model/pytorch_tcts.py) in the current Qlib baseline.
#### Setting1
* Dataset: We use the historical transaction data for 300 stocks on [CSI300](http://www.csindex.com.cn/en/indices/index-detail/000300) from 01/01/2008 to 08/01/2020. We split the data into training (01/01/2008-12/31/2013), validation (01/01/2014-12/31/2015), and test sets (01/01/2016-08/01/2020) based on the transaction time.
* The main tasks <img src="https://latex.codecogs.com/png.latex?T_k" title="T_k" /> refers to forecasting return of stock <img src="https://latex.codecogs.com/png.latex?i" title="i" /> as following,
<div align=center>
<img src="https://latex.codecogs.com/png.image?\dpi{110}&space;r_{i}^{t,k}&space;=&space;\frac{price_i^{t+k}}{price_i^{t+k-1}}-1" title="r_{i}^{t,k} = \frac{price_i^{t+k}}{price_i^{t+k-1}}-1" />
</div>
* Temporally correlated task sets <img src="https://latex.codecogs.com/png.latex?\mathcal{T}_k&space;=&space;\{T_1,&space;T_2,&space;...&space;,&space;T_k\}" title="\mathcal{T}_k = \{T_1, T_2, ... , T_k\}" />, in this paper, <img src="https://latex.codecogs.com/png.latex?\mathcal{T}_3" title="\mathcal{T}_3" />, <img src="https://latex.codecogs.com/png.latex?\mathcal{T}_5" title="\mathcal{T}_5" /> and <img src="https://latex.codecogs.com/png.latex?\mathcal{T}_{10}" title="\mathcal{T}_{10}" /> are used in <img src="https://latex.codecogs.com/png.latex?T_1" title="T_1" />, <img src="https://latex.codecogs.com/png.latex?T_2" title="T_2" />, and <img src="https://latex.codecogs.com/png.latex?T_3" title="T_3" />.
#### Setting2
* Dataset: We use the historical transaction data for 300 stocks on [CSI300](http://www.csindex.com.cn/en/indices/index-detail/000300) from 01/01/2008 to 08/01/2020. We split the data into training (01/01/2008-12/31/2014), validation (01/01/2015-12/31/2016), and test sets (01/01/2017-08/01/2020) based on the transaction time.
* The main tasks <img src="https://latex.codecogs.com/png.latex?T_k" title="T_k" /> refers to forecasting return of stock <img src="https://latex.codecogs.com/png.latex?i" title="i" /> as following,
<div align=center>
<img src="https://latex.codecogs.com/png.image?\dpi{110}&space;r_{i}^{t,k}&space;=&space;\frac{price_i^{t+1+k}}{price_i^{t+1}}-1" title="r_{i}^{t,k} = \frac{price_i^{t+1+k}}{price_i^{t+1}}-1" />
</div>
* In Qlib baseline, <img src="https://latex.codecogs.com/png.latex?\mathcal{T}_3" title="\mathcal{T}_3" />, is used in <img src="https://latex.codecogs.com/png.latex?T_1" title="T_1" />.
### Experimental Result
You can find the experimental result of setting1 in the [paper](http://proceedings.mlr.press/v139/wu21e/wu21e.pdf) and the experimental result of setting2 in this [page](https://github.com/microsoft/qlib/tree/main/examples/benchmarks).Shown in full with attribution under the source's licence. Licence: MIT
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.