Sales Forecasting with ARIMA, Monte Carlo Simulations, and LSTM

In this project, I took on the challenge of predicting future sales using three different forecasting methods: ARIMA, Monte Carlo Simulations, and LSTM. I started by gathering data from two CSV files—one with sales figures and the other with calendar dates. After merging and cleaning the data, including removing outliers, I was ready to dive into the forecasting models.

After running these forecasts, I combined the results into a single DataFrame and compared the accuracy of each model using following metrics: MAE, MSE, RMSE, MAPE, and R². The results showed that Monte Carlo Simulations performed best in terms of mean squared error, while LSTM provided competitive results. ARIMA struggled a bit with the more complex data patterns.

Although the Monte Carlo model proved to be the most accurate, the results obtained in all tests are far from what we might consider "robust." A more in-depth evaluation of the statistical significance of the models is needed, which was not done in this analysis due to its scope. Additionally, improvements can be implemented by considering, for example, the following points:

• Adding more features like promotions or holidays could make the models more accurate. Including such factors might help capture additional trends and patterns.

• The LSTM model could be improved with more in-depth hyperparameter tuning. Experimenting with different configurations, like epochs and batch sizes, could enhance its performance.

• Combining the forecasts from ARIMA, Monte Carlo, and LSTM could lead to better overall accuracy. An ensemble approach might blend the strengths of each model for more robust predictions.

• Using a larger dataset or a longer historical period could provide more insights and improve model performance.

• Setting up a system for real-time forecasting would make the predictions more dynamic, allowing for continuous updates as new data comes in.