ML Ops

See notebook arima_forecast.ipynb in the MLOps repo.

Codebase here: https://github.com/jed-gore/MLOps

A test case to begin to develop a reusable, scalable ARIMA module.

from statsmodels.tsa.statespace.sarimax import SARIMAX

PROCESS:

1. Gather Data
2. If not stationary – apply transformations
3. d = number of times series is differenced
4. List values of p and q
5. Fit every combination in a loop –
6. Select model with lowest AIC
7. Observe residual analysis
8. If uncorrelated residuals … we’re ready for forecasting!

Using Daloopa to pull data for AMZN (you will need to get your own API key):

We isolate Net Sales

And difference it twice to remove the trend and seasonality:

our ARIMA looks ok:

The ARIMA model looks better than naive seasonal:

and the MAPE (% error) is better:

So – this model appears to be useful for forecasting.

Citation: https://www.amazon.com/gp/product/B0BC9W1VBM/ref=ppx_yo_dt_b_search_asin_title?ie=UTF8&psc=1