Master’s Dissertation – How to Enhancing Sales Forecasting with Time Series Analysis?

In my research, I took several powerful machine learning models for a spin, with one goal in mind: making time series forecasting more accurate—especially for sales data that’s impacted by complex time dependencies and seasonal fluctuations. I tested DeepAR, Graph Convolutional Networks (GCN), XGBoost, LightGBM, and Pooling Regression to see which model could really crack the code on sales trends.

The results were pretty surprising! DeepAR not only excelled at capturing long-term dependencies and seasonal changes, but its LSTM network took the forecast accuracy to the next level, leaving the Weighted Root Mean Squared Scaled Error (WRMSSE) in the dust! On the flip side, although GCN showed great promise in modelling the nonlinear relationships in graph-structured data, its performance was less than stellar due to dataset-specific limitations and the heavy computational load.

My research really drove home the point that choosing the right model isn’t a decision to be taken lightly! It’s not just about the data— it’s also about considering the computational resources you’ve got at hand. It’s a bit like cooking: with the right ingredients and tools, you can make a dish that’s truly exceptional. By carefully selecting the models, or even combining a few, you can boost the accuracy of time series forecasting and make sales predictions more reliable than ever.