Forecasting future total sales of ABCD's items based on the historical total sales.

The data is day-to-day total sales of ABCD's items from 21.08.2022 to 01.11.2023. It includes 144 datapoints.
No pre-processing steps were needed for the data, it didn't include any missing or outlier values.

Manual feature engineering was done for the LSTM and Prophet. Lagged sales from 1 to 5 days were added and rolling mean of 5, 10, 15 and 20 days' sales were added. Adding Features did improve Prophet's performance. ARIMA has built-in feature engineering for moving average, lagged features and degree of differencing. The optimal values for these were found with grid search and they were (40, 2, 2) meaning there was 40 lagged observations in the model, degree of differencing was 2 and the moving average windows was 2.

There are many models that can be used for Time-Series forecasting. These include Linear Regression, ARIMA, Random Forest, XGBoost, LSTM, DeepAR and Prophet. In this project all of these were tried except DeepAR. LSTM, Prophet and ARIMA were fine-tuned further. LSTM would need more data. After fine-tuning and feature engineering ARIMA turned out to be the most accurate of these. The ARIMA was fine-tuned using grid search. It significantly improved performance.
Prophet: 55.80 MAE (and odd looking prediction) without feature engineering or 23.71 with feature engineering
LSTM: 39.28 MAE
ARIMA: 21.93 MAE

For evaluation Mean Absolute Error (MAE) was used. It is calculated as the average of the absolute differences between the predicted and actual values. In other words, it measures the average magnitude of the errors in a set of predictions, without considering their direction. The lower the MAE, the better the model's performance.

The Mean Absolute Error (MAE) of the ARIMA model was 21.93, this means the model's prediction from the actual sales in the test set was off by 21.93 points on average. The model also gives a 95% confidence interval of the predictions. This means that there's a 95% chance the true value lies between these intervals. As you can see in the plot, the interval gets bigger as the predicted timeframe gets longer.

An API endpoint could be created that takes input data and returns the model's prediction.

Once the model is deployed, it is important to monitor its performance. This can be done by tracking the MAE to ensure that the model is performing as expected.

Regular maintenance of the model is important to ensure that it continues to perform well. This includes e.g. retraining with new data and updating model parameters.

An alerting system should be set up which will notify the team if the model's performance drops below a certain threshold.

A log of all the requests and predictions made by the model should be kept, this will be helpful in monitoring and debugging the model.

For LSTM, ARIMA and Prophet you can see the development in different .ipynb -files and the best ARIMA model is sales_model.sav. You can use that for the predictions.