Advanced Demand Forecasting Part 2: A technical look
As I mentioned in Part 1, Advanced Demand Forecasting is not a one-size-fits-all solution. It is uniquely tailored for each business and requires about 6+ months to deploy. But how does it work?
Advanced Demand Forecasting leverages two techniques:
- Auto-Regressive Integrated Moving Average (ARIMA) or Error Trend Seasonality (ETS) analysis to produce the forecasts
- Residual regression analysis to determine what factors are driving demand beyond seasonal changes
When implementing Advanced Demand Forecasting, we typically will leverage ARIMA when adaptability to change is crucial and data exits at constant intervals. If the data follows unique seasonal patterns and does not have consistent observation intervals, we will typically leverage ETS.
Once the forecasting model is complete, regression analysis is performed on the model’s residuals using customer, product, and sales data to determine what factors move demand independent of yearly seasonality. For example, this sophisticated approach can show how an inch of snow above normal levels, or a dollar price in adjustment, can impact demand.
The techniques mentioned above all use the latest open-source Python packages from Statsmodels, Prophet, and sci-kit-learn. We can also run the modeling using Spark ML packages, which allow us to run our modeling on massive datasets using the latest parallelized computing techniques with resources like Azure Databricks.
The data scientists at Neal Analytics use these techniques to look at factors that affect each business and develop forecasts to discover which ones have the most significant impact. By looking at how demand has changed in the past and using other company or market data from previous periods, Advanced Demand Forecasting can provide businesses much more accurate and, more importantly, dynamic models.
Interested in learning more about Advanced Demand Forecasting? Check out part 3 of this three-part blog series.