Supervised Machine Learning (SML) algorithms, such as classification and regression trees, linear regression and time series, have proven themselves to be the backbone of forecasting components needed for Enterprise Revenue Management and Pricing System (ERMPS) solutions. Forecasting techniques that use SML algorithms make predictions by “training” the SML algorithm against a historical set of predictors and target variables.
In my previous blog, I explored volatile data conditions, forecast error, model misspecifications and environmental disruptions as the typical unknowns that need to be considered when determining the solution for an ERMPS. Below are key considerations that can be used to tailor ERMPS forecast solutions to better adapt to environmental disruptions and to protect against the risk of model misspecifications.
1) Determine appropriate methods to help a model learn. There are two methods that can be used in conjunction with each other to help models learn: historical forecasting components that are used to protect against the risk of model misspecifications, and current forecasting components that are used to adapt to current environmental disruptions.
Methods used for the historical forecasting component include:
- Classification and Regression Tree (CART) Estimation, which identifies key segmentation attributes for the forecast.
- Normalization, which calibrates historical data against key non-price market factors to determine business-as-usual forecasting state.
- Pick-Best Forecasting, which calibrates multiple forecasting models based on out-of-sample forecast accuracy metrics to identify the best functional form to be used for future forecasts.
Methods used for current forecasting component include:
- Price Sensitive Adjustment, which helps to adjust forecast trends based on the difference between current price and business-as-usual price.
- Pace Adjustment, which normalizes forecast trends based on shifts in recent trends and/or year-over-year trends.
- Adaptive Gain, which regulates forecast level based on sudden and significant shifts in recent observed values.
- Classification and Regression (CART) Prediction, which re-calculates the value of the forecast target variable based on CART estimates and changes in the set of predictors over time.
2) Strike the right balance. Decision-makers must toe the line between system stability and the forecasting solution’s ability to adapt to recent changes in the history/projected state. To do this, they must answer questions such as:
- What is the appropriate frequency of recalibrating the historical and current model (quarterly, monthly, daily, hourly)?
- What is the user criteria for a volatile forecast?
The willingness to accept a forecasting solution that is more adaptive is hindered if volatility criteria are not met and frequent user interaction are required. In addition, it’s essential to forecast as close as possible to the level at which users manage the forecast and distribute down.
3) Establish control parameters and business rules. One of the most critical elements for determining the ERMPS forecasting solution is the identification of a set of control parameters and business rules that artfully blend the science behind adaptive forecasting techniques with business reasonability checks. In order to tailor the demand forecast to fit specific needs, gain user acceptance and ensure it aligns with corporate objectives, decision makers must answer questions such as:
- What is the minimum or maximum demand change from current state that triggers a demand re-forecast?
- What is the maximum demand change from current state that users are willing to accept?
- What is the appropriate combination and sequence of models, control parameters and business rules?
Control parameters can be used to either control which forecast functionality gets turned on/off or tune models algorithms to better align with particular business / data conditions. Business rules can be used to correct abnormal data inputs (pre-process), to enforce specific model behavior (constraints), or to create a safety net that ensures the results that are shared with users are business reasonable (post-process). It’s important to store the raw model results before applying business rules, and have on-going monitoring processes that quantify the frequency and magnitude of control parameters and business rules changing raw model results.
Supervised Machine Learning (SML) Algorithms are the foundation for constructing adaptive forecasting solutions. To capture the full potential of your adaptive SML forecasting solution, consider the appropriate combination of historical and current components that help the model learn, strike the right balance between forecast stability and forecast reactiveness, establishing control parameters and business rules that ensure the forecast is in line with corporate objectives and well-positioned to gain user acceptance.