DemandCaster currently uses standard statistical methods like exponential smoothing, the Croston Method, bootstrapping, and Syntetos-Boylan Approximation (SBA) to forecast demand. These methods are quite accurate and robust. The process generates forecasts very quickly. However, there are many instances where the results are not as accurate as they could be.

Machine learning forecasting leverages a different process to generate time series forecasts. Machine learning methods automatically evaluates the data to identify the key attributes and then select the best algorithms for forecasting. From there it trains and optimizes the model.

Adding additional related data sets such as average sales price, holidays, events, IHS, and POS to name a few further improve accuracy by correlating the shipment history against external influencing events. Please note that the related time series option will be released soon.

For machine learning to be more accurate, it is helpful to have enough history for the ML engine to learn and apply the appropriate model. Advanced deep learning algorithms require a particularly large set of historical data - ideally a 4:1 ratio between the number of periods of history versus periods of forecast. For example, 2 or more years of history is required to forecast the next 6 months of demand.

Insufficient historical data leads to fewer input data points, which may not be enough for machine learning to accurately train the predictor.

The current statistical forecasting process allows for hierarchical forecasting.

Machine Learning Forecast Manager instead forecasts at the PCL. This is because all the related data is captured at the customer, location, and product detail level which is pulled from the sales record.

Machine Learning Forecast Manager is an add-on to Market Forecast Manager. The solution provides customers the option to select the number of forecasts they would like to generate on a monthly basis. This volume is defined as:

Number of Forecast Entities (PCL's) multiplied by the number of forward periods that are forecasted.

For example, if 10 PCL's are being forecasted out 52 weeks once a month, that is equal to a forecast volume of 520 forecasts.

Related time series data refers to data that is related to the time series being forecasted. For example, if you are forecasting demand for a particular product, related time series data might include factors such as advertising spend, promotional activity, seasonality, or economic indicators.

By incorporating related time series data into the forecasting model, we can improve the accuracy of our predictions. This is because related time series data can capture important factors that may influence the primary time series being forecasted. These factors may not be immediately apparent or may not have a direct relationship with the primary time series, but they can still impact its behavior over time.

At Demand Caster, we have introduced a new feature in our Machine Learning add-on called "Price as Related Time Series". This feature allows users to include price data as a related time series in their forecasting models. Price is a critical factor in many industries and can have a significant impact on sales and demand. By including price as a related time series, our users can improve the accuracy of their demand forecasts.

Machine learning is more accurate in short term forecasts - up to 52 weeks in the future with 4 years of history. Beyond that time frame, the forecast will be driven by DemandCaster's built in forecasting algorithms. Since the solution combines two forecast types, a new line is added to the demand plan called "Blended Forecast."

Blended forecast is the time frame where the Machine Learning Forecast automatically overrides the DemandCaster forecast. to populate the plan line. From there, users may further edit the plan line as needed. All accuracy metrics are then generated via the Blended Forecast line.

The image below illustrates how the ML Forecast overrides the Statistical Forecast to establish the non-edited Plan value.

Below is a simpler illustration of how the override works. The first 6 forecast periods are handled by the ML Forecast. Starting in period 7, the DC Stat Forecast takes over. If a new stat forecast is run at the start of the next period, the period that was previously filled with the DC Forecast will be replaced by the ML Forecast.