2022美国大学生数学建模竞赛C题摘要及总结

Posted 2022-03-08 ZSYL

MCM Problem C：贸易策略

• 问题重述
• 论文摘要
• 解决方案
• 个人总结
• 免费资料

问题重述

给定6年黄金和比特币的价格数据，起始资产1000美元，在开盘日可以选择买、卖和保持策略，注意比特币每天都开盘，黄金不是每天都开盘，黄金和比特币的交易手续费分别是1%和2%的本金。

• 问题一：给出决策算法使得6年内收益最大化并且预测最大收益是多少，而且每天的决策只能利用前面和当天的数据。
• 问题二：证明你模型的有效性
• 问题三：当交易手续费变化时，你决策模型的鲁棒性怎么样。
• 问题四：写封信，介绍并推荐你的模型。

论文摘要

Optimal trade strategy based on RNN-LSTM deep learning method

Summary

In this paper, by incorporating the financial time series data dependencies and the local correlation characteristics of the time series data of the two financial markets of gold and Bitcoin into the same model, the LSTM neural network price prediction model is constructed and the parameters are optimized to make it more accurate to predict gold, Bitcoin price trend. At the same time, using the integrated empirical mode decomposition and run-length determination method, the time series data of gold and bitcoin are divided into cycles, and the LSTM neural network model is trained and verified, and finally the gold bitcoin price prediction model is obtained.

For problem 1, firstly, we adopted three different LSTM neural network structures and compared them and found that the bidirectional LSTM neural network model is the most suitable one for predicting gold prices, and the single-layer LSTM neural network model is the most suitable one for Bitcoin price prediction. Secondly, combined with model 2, we calculated the expected value of gold and bitcoin income on the next trading day by day, and determined the trading strategy of the day according to the calculation results and recent trading behavior, and finally calculated that the final income on October 9, 2021 is about $45083.3747, seen in Appendix 2 for the strategy.

For problem 2, two submodels are established. Combining the internal relationship of each subsystem with the price trend of gold bitcoin and the characteristics of the financial environment being easily affected by emergencies, the prediction accuracy of the model is verified by calculating the difference between the predicted price of gold bitcoin and the real price, and the final result is obtained. The mean absolute error and the root mean square error of the model’s prediction of gold and bitcoin prices. We then use the historical trends of gold and bitcoin prices to reduce the risk of errors in the forecasting model.

For problem 3, we consider the impact of the transaction commission and the proportion of a single transaction amount on the overall strategy, calculate the final income value generated by taking different proportions, and draw images for visual display, and finally come to a conclusion. The overall sensitivity of the model is less affected by the transaction cost of gold, and more affected by the transaction cost of Bitcoin.

Finally, we communicate with traders to share and promote our price prediction model and investment strategy model. By analyzing the advantages, disadvantages and applicable conditions of the model, traders can make better use of our trading strategies and obtain higher profits.

Keywords: deep learning; LSTM neural network; portfolio theory; optimal trade strategy

解决方案

• 数据探索：数据探索发现，黄金存在缺失值，双休日和节假日不交易，6年内年平均收益总体呈现上升趋势，月平均收益的趋势多样性。
• 预测模型：基于LSTM做时序预测，最终训练出两个模型，达到的效果是给定前30天的价格可以预测后一天的价格。(主要是数据预处理(缺失值处理、数据归一化、数据划分、训练集/测试集构建)，模型训练，超参调节(seq_len、epochs)，模型评估(均方差损失、rms)，模型保存与使用)
• 分类决策模型：根据预测模型预测出后一天的价格，然后计算增长率，根据增长率选择对应的策略。
• 我们总共把决策分为六类：黄金保持美元全部换成比特币，黄金保持比特币全部换成美元，全保持，全换成黄金，全换成比特币，全换成美元。本质是一种极端的贪心决策算法。分类条件是在开盘的前提下，黄金和比特币哪种的增长率大(趋势好)，并且是否远远大于手续费，远远大于引入两组参数来控制阈值。通过调节阈值的大小，使得模型目光更长远。这样我们的模型更具有远见。(起初没有考虑到用阈值，而是只要大于手续费就行，导致收益率很低)
• 分类决策模型构建完成之后，通过调节交换手续费来看模型的泛化能力，最终发现我们的模型随着手续费增加总体收益率呈现下降趋势但是没有亏损现象。

个人总结

• 数据探索和数据预处理，增加了对数据的认识和处理能力，比如常见的缺失值处理，数据归一化，数据划分，数据集构建等原理和具体实现
• 从零构建深度学习预测模型，熟悉了总体流程，对超参数的调节有一定的认识和理解
• 遇到贪心决策算法不能得到很好的收益率，然后我们研究决策过程，通过讨论发现，可能是贪心过于极端，有种目光短浅的方法，所以我们想能不能让模型把眼光放长远些，最终通过阈值来减少目光短浅，使的模型的眼光更长远。
• 检索能力尤其重要，很多时候想要实现的功能，不知道怎么写代码，其实你只需要相信一点，大多数基本功能常见问题可以通过搜索引擎解决，函数api可以通过API文档和数据测试理解，通过文献调研也可以看看哪种模型可能比较适合解决这类问题。
• 英语阅读和写作能力尤其重要，题目和大多数参考资料是英文的，最终的论文是英文的，虽然有翻译软件辅助，但是翻译软件翻译很长的中文，很难翻译出地道的句子。所以翻译软件只能辅助查单词短语，不能成篇翻译，需要一定的英文阅读和写作能力
• 论文排版也很重要，如果有word模板的一定要拿来直接使用，会方便很多，节省时间。
• 注意比赛的规则，以及论文格式等要求。
• 感谢队友的精诚合作！

免费资料

如果需要国赛论文模板、软件、优秀论文、参考文献等，请关注微信公众号或者私信给我。

---

加油!

感谢!

努力!