Week 2: ARIMA Models

Johnny Y -

This week, I focused on building and optimizing my baseline ARIMA (autoregressive integrated moving average) model. I first built an ARIMA model to predict continuous prices (which ARIMA tends to perform better on). I found the parameters using the Augmented Dickey-Fuller (ADF), Autocorrelation Function (ACF), and Partial Autocorrelation Function (PACF) tests, and used Dogecoin data for the testing and training. I played around with different time frames to see how they affected accuracy.

I then built a separate ARIMA model to predict only whether the price would go up or down on a given day (since that is closer to what sentiment analysis does best) and measured the accuracy using percentage returns as data. With an 80%-20% train-test split on 12 months of data and parameters (p, d, q) = (25, 0, 25), the model achieved a 42.47% accuracy – less than a random coin flip. This would appear to demonstrate that ARIMA perhaps isn’t the best baseline for up/down predictions, though certainly the huge swings in the crypto market (e.g. 2.5x in 1 day because of the election) don’t help. Debugging my code took quite a bit of time, and for some time frames/parameter choices the model seems to have uniform recommendations (i.e. predicting “up” every day or “down” every day); I still haven’t figured that out yet. Another challenge is that manual entry is only weekly, so I really only have 1 years’ worth of data. I will be meeting with my site advisor to discuss the results.

It should be noted that past papers generally used ARIMA as a baseline for predicting continuous prices instead of up/down, so I don’t anticipate the lack of accuracy of ARIMA for up/down movement being an issue, though I will discuss this with my site advisor.

More Posts

Comments:

All viewpoints are welcome but profane, threatening, disrespectful, or harassing comments will not be tolerated and are subject to moderation up to, and including, full deletion.

    tesla_l
    What is a train test? Are there other baseline models you might consider using instead?
    February 22, 2025 at 6:11 pm - Reply
    rob_lee
    Inverse -ARIMA?
    February 24, 2025 at 1:18 pm - Reply
    makeen_s
    Hi Johnny. It sounds concerning that you only have around a years worth of data. Would it make sense to investigate a broader scope of coins to get more data?
    February 24, 2025 at 3:26 pm - Reply
    austin_l
    How would you increase the accuracy of your model? Would you want to switch or manipulate something in your existing one?
    February 24, 2025 at 4:25 pm - Reply
    Anonymous
    If the ARIMA model is less accurate in predicting up/down movement, can it be used in conjunction with another model to accurately predict both up/down and continuous prices?
    February 24, 2025 at 9:20 pm - Reply
    johnny_y
    Thank you for your questions, Tesla! By 80%-20% train-test split I mean 80% of the data was used for training and 20% of the data was used for testing. Other potential baselines I can think of are random choice, tweet/post/comment volume, and random forest classifier, though I'm sure my site mentor can suggest more alternatives.
    February 25, 2025 at 11:00 am - Reply
    johnny_y
    Mr. Lee, could you elaborate on what you mean by inverse ARIMA? Converting the data back before calculating error?
    February 25, 2025 at 11:22 am - Reply
    johnny_y
    Thank you for your question, Makeen! Based on past papers, 1 year of data should be enough (especially since I have 3 coins), though it's possible I might need to expand the scope.
    February 25, 2025 at 11:24 am - Reply
    johnny_y
    Thank you for your question, Austin! I'll be meeting with my mentor later this week to see how I can improve the model's performance. I may need to combine it with machine learning like LSTM or a random forest classifier.
    February 25, 2025 at 11:27 am - Reply
    johnny_y
    Yes, another model might be better suited for up/down prediction. I don't anticipate issues running that model alongside ARIMA.
    February 25, 2025 at 11:29 am - Reply

Leave a Reply

Your email address will not be published. Required fields are marked *