A Novel Hybridized Forecasting Technique Utilizing ARIMA and Large Language Models

Usage

1. This project requires Ollama, so ensure it is installed and running.
2. For web scraping, you need to have Mozilla Firefox installed.
3. Install requirements by running pip install -r requirements.txt in the directory of this project. It is recommended you do this in a virtual environment.
4. To replicate our tests, first delete the Data, Results, and Responses folders.
5. Run main.py to perform the experiments. Sometimes, during the web scraping Selenium using Mozilla Firefox may hang, and not move onto the next news article. In this case, simply restart the script, and it will continue from where it left off.
   1. -f or --forecast - Number of weeks to forecast. Defaults to twelve.
   2. -w or --width - The width of the figures. Defaults to eight.
   3. -t or --height - The height of the figures. Defaults to 3.45.
   4. -d or --decimals - The number of decimal spaces. Defaults to two.
   5. -a or --alpha - The alpha factor for the desired confidence level which by default is 95%.
   6. -c or --clamp - By how much should forecast values be clamped around the baseline prediction. Defaults to one hundred.
   7. -l or --latest - Up to how many latest weeks of data should we keep. Defaults to zero meaning keep all data.
6. Under the Data folder, you will see the summaries of the news articles produced by the large language model. Under the Results folder, you will see the results charts and plots.
   1. Actual.csv - The actual COVID-19 hospitalizations to occur over the next given weeks from a given week.
   2. Baseline.csv, Unmasked.csv, and Masked.csv - The baseline model and full model predictions of how many COVID-19 hospitalizations to occur over the next given weeks from a given week.
   3. Difference Baseline.csv and Difference Full Model.csv - The difference between the actual results and each of the model results.
   4. Success Rate.csv - The success rate of each model, where success was determined if a forecast met or exceeded the actual amounts of hospitalizations to occur over a period.
   5. Average Difference.csv - The average difference each model had from the actual amounts of hospitalizations to occur over a given period.
   6. Total Failures.csv - The total failures which occurred for each model over a forecasting period.
   7. Total Excess.csv - The total excess which occurred for each model over a forecasting period.
   8. WIS.csv - The weighted interval scores.
   9. MAE.csv - The mean absolute errors.
   10. Coverage.csv - Coverage of values falling within the alpha prediction intervals.
7. Terms.txt included all terms for COVID-19 which should be masked.
8. Trusted.txt includes all trusted publishers.
   1. To list all untrusted publishers in all the summarized news articles, run publishers.py. Passing in either -t or --trusted will list all trusted publishers in all the summarized news articles.

Data

The data used for this experiment is taken from Covid Timeline Canada.