[论文解读] Bitcoin Price Prediction: An ARIMA Approach
本文使用三年数据集评估了ARIMA模型在短期比特币价格预测中的表现,发现在稳定子时段内有效,但在捕捉剧烈波动方面存在局限,尤其是在2017年价格飙升期间。主要贡献在于识别出ARIMA性能在很大程度上取决于窗口大小和参数选择,随着窗口大小从2天增至9天,最优模型将均方误差从118,000降低至16,000。
Bitcoin is considered the most valuable currency in the world. Besides being highly valuable, its value has also experienced a steep increase, from around 1 dollar in 2010 to around 18000 in 2017. Then, in recent years, it has attracted considerable attention in a diverse set of fields, including economics and computer science. The former mainly focuses on studying how it affects the market, determining reasons behinds its price fluctuations, and predicting its future prices. The latter mainly focuses on its vulnerabilities, scalability, and other techno-crypto-economic issues. Here, we aim at revealing the usefulness of traditional autoregressive integrative moving average (ARIMA) model in predicting the future value of bitcoin by analyzing the price time series in a 3-years-long time period. On the one hand, our empirical studies reveal that this simple scheme is efficient in sub-periods in which the behavior of the time-series is almost unchanged, especially when it is used for short-term prediction, e.g. 1-day. On the other hand, when we try to train the ARIMA model to a 3-years-long period, during which the bitcoin price has experienced different behaviors, or when we try to use it for a long-term prediction, we observe that it introduces large prediction errors. Especially, the ARIMA model is unable to capture the sharp fluctuations in the price, e.g. the volatility at the end of 2017. Then, it calls for more features to be extracted and used along with the price for a more accurate prediction of the price. We have further investigated the bitcoin price prediction using an ARIMA model, trained over a large dataset, and a limited test window of the bitcoin price, with length $w$, as inputs. Our study sheds lights on the interaction of the prediction accuracy, choice of ($p,q,d$), and window size $w$.
研究动机与目标
- 使用历史收盘价评估ARIMA模型在预测比特币短期价格变动方面的有效性。
- 研究ARIMA参数(p, d, q)和训练窗口大小的选择对预测准确率的影响。
- 识别在比特币高波动性和非平稳行为背景下ARIMA仍具可行性的条件。
- 比较基于RSS最小化与MSE最小化的ARIMA模型选择策略,以提升预测性能。
- 评估时间窗口位置与长度对不同市场周期下预测误差的影响。
提出的方法
- 通过取对数变换和一阶差分对比特币价格数据进行预处理,以实现平稳性,经ADF检验验证,p值接近零。
- 在p ∈ {0,…,9},q ∈ {0,…,9},d ∈ {0,…,2}范围内进行网格搜索,以识别最优ARIMA配置,通过残差平方和(RSS)和均方误差(MSE)评估模型。
- 采用两种模型选择策略:一种在拟合过程中最小化RSS,另一种在预测过程中最小化MSE,以避免过拟合。
- 通过改变时间窗口大小(w = 2至9)进行实验,并在三年跨度(2015–2018)内随机选择窗口进行预测测试。
- 分析预测误差随时间的分布,识别出高波动期(如2017年底)导致MSE显著上升。
- 结果以两种设置报告:全数据集训练,测试窗口覆盖整个时间跨度(表I);全数据集训练,但仅在前半段(2015–2017)进行测试(表II)。
实验结果
研究问题
- RQ1当使用包含不同市场行为的三年数据集进行训练时,ARIMA模型在短周期(1天)内对比特币价格的预测效果如何?
- RQ2ARIMA参数选择(p, d, q)对预测准确率有何影响,特别是在RSS最小化与MSE最小化之间权衡时?
- RQ3训练窗口长度(w)如何影响比特币价格预测的均方误差(MSE)?
- RQ4测试窗口在时间序列中的位置如何影响预测误差,特别是在2017年底等波动剧烈时期?
- RQ5ARIMA模型能否有效捕捉如2017年价格飙升及随后调整等剧烈价格波动,还是其假设从根本上限制了其表现?
主要发现
- ARIMA(8,8,1)模型将RSS最小化至0.002,但其MSE接近最优模型的100倍,表明对训练数据存在过拟合。
- 当使用9天窗口时,ARIMA(4,1,1)模型实现了最低MSE(16,000),显著优于RSS优化模型。
- 在2017年价格飙升与调整期间,预测误差最高,MSE峰值超过100,000,凸显ARIMA在捕捉极端波动方面的无能。
- 当在全数据集上训练但仅在前半段(2015–2017)进行测试时,平均MSE降至167,表明在稳定、非波动子时段测试时模型性能更优。
- MSE最小化模型(如w=9时的ARIMA(4,1,1))的RSS高于RSS最小化模型,表明拟合准确率与预测性能之间存在权衡。
- 随着窗口大小从2天增至9天,平均MSE从118,000降至16,000,表明在波动市场中,更长的窗口可提升预测稳定性。
更好的研究,从现在开始
从阅读论文到最终审阅,大幅缩短您的研究时间。
无需绑定信用卡
本解读由 AI 生成,并经人工编辑审核。