12 Forecasting Methods in AIF

This chapter discusses the forecast methodologies available in Oracle’s Next Generation AI Foundation (AIF). The intent of this chapter is to highlight forecast algorithms that are available with Oracle’s Retail & Analytics Planning (RAP) cloud services within AIF. RAP solutions consist of Assortment Planning (AP), Assortment Space Optimization (ASO), Inventory Planning Optimization (IPO), LifeCycle Pricing Optimization (LPO), Merchandise Financial Planning (MFP), and Retail Insights (RI). To facilitate the retailer’s planning and analytics needs, RAP solutions may be deployed with the forecasting capabilities of AIF. The utilization of forecasting methodologies discussed in the chapter will differ based on the cloud services implemented and the unique requirements of each retailer’s business.

Some Forecasting Concepts

In the Next Generation architecture, the AI Foundation supports all the RAP cloud services by providing a single data point of entry to multiple cloud services. From this data, the Forecasting Engine fuels different forecasts at the different levels required by the different processes of the different cloud services. MFP and AP utilize the base functionality of the Forecast Engine. IPO utilizes the Forecasting Engine and the Forecasting Workflow. IPO offers enhanced forecasting review and workflow management that maintain and adjust forecasts at the item level. The other solutions of ASO, LPO, and RI may be configured to utilize the

AIF forecasts as best serves the retailers business objectives. AIF forecasts enhance retailer supply chain’s planning, allocation, and replenishment processes for a more profitable and customer-oriented approach. This unconstrained demand may be sent to execution and planning solutions to be transformed into revenue-based metrics and strategies. Demand forecast provides a single unified forecast with minimal human intervention. The Forecasting Engine generates accurate forecasts of demand as automatically as possible by using a variety of predictive techniques.

The AIF Forecasting Engine utilizes several concepts when forecasting for planning solutions. These concepts are integral parts of the various forecasting methods available in RAP cloud solutions.

Exponential Smoothing

Exponential Smoothing (ES) evaluates the basic demand patterns of Level, Trend, and Seasonality and projects these into the future. They are labelled smoothing since they use weighted averages on historic data, and the weighting decays exponentially as we go back in time: more recent data is weighted more heavily than past data.

Figure 12-1 Demand Patterns of Level, Trend, Seasonality

Automatic Method Selection

Multiple forecasting methods are valuable if the appropriate model can be selected in an accurate and efficient manner. Oracle Retail has developed different meta-methods that can automatically select the best method among competing models. These models include level information, level and trend information, and level, trend, and seasonality information. The optimal smoothing parameters for each model form are determined automatically (that is, greater smoothing is applied to noisier data). The final selection between the resulting models is made according to a performance criterion that involves a trade-off between the model's fit over the historic data and its complexity. The amount of available historic information can affect the complexity of the model. Automatic Exponential Smoothing (Auto ES) and Seasonal Exponential Smoothing (Seasonal ES) chooses between competing models according to performance criterion, looking at the tradeoff between history and its complexity.

Seasonal Regression Analysis

Regression analysis is another standard technique used in prediction. Regression uses a least-squares estimator to fit a model of predictor variables to another set of target variables. Seasonal regression is Oracle Retail’s extension to the standard seasonal forecast method for application between one and two years of history. This is included in the Automatic Exponential Smoothing (Auto ES) and Seasonal Exponential Smoothing (Seasonal ES) method.

Bayesian Analysis

Bayesian Analysis is effective where no previous sales history exists, such as with new items or items with short lifecycles. It requires a custom plan to serve as the initial sales plan. It may be configured to lend more weight to recent history as it accrues. Once enough history exists, standard forecasting methods must be used, such as Exponential Smoothing. This enables a plan with current data to create a new forecast.

Forecasting Methologies

From the various forecasting concepts detailed above, the following forecasting methodologies are created. These forecast methods are available to the user when creating a forecast run type within the AIF Control and Tactical Center -> Strategy and Policy Management -> Manage Forecast Configurations screen. Different methods have different data needs. Refer to the Data Requirements for Various Forecast Methods section from the AIFCS Implementation guide for more information.

• Automatic Exponential Smoothing

• Sales & Promo

• Life Cycle

• Hierarchical Bayesian

Automatic Exponential Smoothing Method

The primary goal for a demand forecasting solution is automation. The Automatic Exponential Smoothing method (Auto ES) automatically selects an exponential smoothing model to a time series. Auto ES is a proprietary forecasting algorithm that compares the varying ES methods against each other for the best fit while being the least complex.

Auto ES considers a collection of six candidate models:

1. Simple Exponential Smoothing (Simple ES):

   • Simple ES ignores seasonality and trend features in the demand data.

   • Simple ES is effective when there are short time horizons in which less than one year of historic demand data is available or data is untrended and unseasonal.

   • Simple ES produces a flat line (Level only) forecast.

2. Holt Exponential Smoothing (Holt ES):

   • Holt ES captures trends but is non-seasonal.

   • Holt ES treats data as a one-dimensional trend (up or down).

   • Holt ES provides forecast estimates by combining an estimated trend for the forecast horizon and the smoothed level at the end of the series.

   • Holt ES is effective for short or long horizons and provides a damping effect for long horizons.

3. Additive Winters Exponential Smoothing (Additive Winters ES):

   • Additive Winters ES extracts seasonal indexes assumed to have additive effects on the de-seasonalized series. Effects are applied additively to levels and trends.

   • Additive Winters ES requires sufficient data of at least two years of history.

   • Additive Winters ES controls the smoothing of the components of level, trend, and seasonality.

   • Additive Winters ES is effective when there are zeros in the history.

4. Multiplicative Winters Exponential Smoothing (Multiplicative Winters ES):

   • Multiplicative Winters ES extracts seasonal indexes assumed to have multiplicative effects on the de-seasonalized series. Effects are applied multiplicatively to levels and trends.

   • Multiplicative Winters ES requires sufficient data of at least two years of history.

   • Multiplicative Winters ES controls the smoothing of the components of level, trend, and seasonality.

   • Multiplicative Winters ES does not work when there are zeros in the history.

5. Seasonal Regression:

   • Seasonal Regression is for sales forecasts based entirely on sales from the same time period last year. Forecasting using only last year sales involves simple calculations and often outperforms more sophisticated seasonal models. –

   • Seasonal Regression performs best when dealing with highly seasonal sales data with a short sales history.

   • Seasonal regression is designed to address retailer needs when sales history is between 56 to 104 weeks.

   • Seasonal Regression uses simple linear regression with last year’s sales as the predictor variable and this year’s sales as the target variable.

   • If there are significant shifts from year to year, Seasonal Regression will learn the shift and appropriately weigh last year’s data, keeping the same shape but adjusting the scale.

6. Bayesian:

   • The Bayesian method combines historic sales data with sales plan data and is effective for new products with little or no historic data. The sales plan provides expert knowledge. The sales plan defines the shape of the selling profile or lifecycle and the scale of the total quantity expected to be sold over the plan's duration. The Bayesian method merges a customer’s sales plans with any available historical sales.

   • The Bayesian method starts with a seasonal sales plan and evaluates the following conditions to calculate the forecast:

     • When no sales history exists, the forecast equals the sales plan.

     • As sales data becomes available, the forecast adjusts, and the scale becomes a weighted average reflecting known history.

     • Once sufficient sales data is collected, it can be switched to one of the different standard forecasting methods.

During configuration, an implementer may determine if only Seasonal Exponential Smoothing (Seasonal ES) methodologies are appropriate and remove access to Simple ES and Holt ES methods.

Sales & Promo Method

This method is useful for retailers wanting to account for outliers, out-of-stock, and promotion events. This is a causal approach and goes through multiple stages to generate the final forecast. Typically, the data requirement for this method is at least two years of sales history. This is among the best methods in AIF that can produce reliable forecasts at the lowest levels of merchandise and location hierarchies with only the minimum requirement of sales data. The stages for Sales & Promo method:

1. Preprocessing corrects bad data points that represent unusual sales behavior, so that the demand forecast does not replicate undesirable patterns. Possible causes of history adjustments include out of stocks, outliers, or events. During preprocessing, the quality of data is analyzed, and faulty data points are corrected before further calculations. This provides pre-processed data to produce reliable demand parameters for the next process steps.

2. Promo Estimation consists of heavy data mining of the relationships between historical sales (sales lift) and promotions (different promotion-relevant features). The two available models to quantify the impact of promotions on demand are:

   • Decision Trees:

     • Tree-based Machine Learning classifier model that predicts an outcome based on several input features.

   • Generalized Linear Model:

     • Regression-based Statistical model that predicts an outcome based on the values of another set of input variables.

3. Seasonality Estimation utilizes various machine learning techniques to understand the time series in which data experiences regular and predictable changes that recur every calendar year. This reflects a sales pattern that repeats over a one-year period based on the time of year (season). The estimation of seasonality occurs at different Escalation levels which are aggregated sales at different merchandise and location hierarchy levels. The four options for estimating seasonality are:

   • Additive Winters Exponential Smoothing (Additive Winters ES):

     • Additive Winters ES extracts seasonal indexes assumed to have additive effects on the de-seasonalized series. Effects are applied additively to levels and trends.

     • Additive Winters ES requires sufficient data of at least two years of history.

     • Additive Winters ES is effective when there are zeros in the history.

   • Multiplicative Winters Exponential Smoothing (Multiplicative Winters ES):

     • Multiplicative Winters ES extracts seasonal indexes assumed to have multiplicative effects on the de-seasonalized series. Effects are applied multiplicatively to levels and trends.

     • Multiplicative Winters ES requires sufficient data of at least two years of history.

     • Multiplicative Winters ES does not work when there are zeros in the history.

   • Auto Seasonal:

     • Selects the best method based on the inverse of Root Mean Square Error of Additive Winters ES or Multiplicative Winters ES.

   • Blend Seasonal:

     • Utilizes the inverse of Root Mean Square Error as the weights for the two methods of Additive Winters ES and Multiplicative Winters ES and combines the weighted seasonal indexes.

4. Base Demand Estimation calculates the average demand signal after the cleansing of data and removal of promotional and seasonality effects. This is generated at the same merchandise and location hierarchy level as the required forecast. The six options for estimating base demand are:

   • Simple Exponential Smoothing (Simple ES):

     • Simple ES ignores seasonality and trend features in the demand data.

     • Simple ES produces a flat line (Level only) estimate.

   • Holt Exponential Smoothing (Holt ES):

     • Holt ES captures trends but is non-seasonal.

     • Holt ES treats data as a one-dimensional trend (up or down).

     • Holt ES provides estimates by combining an estimated trend for the forecast horizon and the smoothed level at the end of the series.

     • Holt ES provides a damping effect for long horizons.

   • Moving Average:

     • Moving Average is calculated as a simple moving average estimate from past time frames.

     • Moving Average gets priority over Simple ES and Holt ES when the data is too noisy.

   • Poisson:

     • This helps in estimating for low sellers assuming a Poisson distribution on the sales history.

   • Pick Best:

     • System determines which method (Simple ES, Holt ES, Moving Average, or Poisson Base Demand) provides the best fit based on the inverse of Root Mean Square Error.

   • Blend:

     • System generates estimates from various methods (Simple ES, Holt ES, Moving Average, or Poisson Base Demand) and uses the inverse of Root Mean Square Error as weights to combine them.

5. Forecast Generation calculates the demand forecast combining the output of all previous steps. This forecast multiplies base demand, seasonality, and promotional lifts. In this manner, changes in seasonality and promotional effects can be evaluated and combined with the de-seasonalized base demand.

Life Cycle Method

This method is useful for retailers wanting to account for price elasticities and markdowns. This is a causal approach and goes through multiple stages to generate the final forecast. Typically, the data requirement for this method is more than one year of sales, inventory, receipts, and prices. This is among the best methods in AIF that can produce reliable forecasts at the lowest levels of merchandise and location hierarchies accounting for prices and markdowns but with a larger breadth of data needs.

The stages for Life Cycle method:

1. Preprocessing filters bad data points that represent unusual sales behavior, low inventory, or irregular prices so that the demand forecast does not replicate undesirable patterns. The filtering of historical data defines a subset of data to produce reliable demand parameters for the next process steps.

2. Event Lift Estimation involves heavy data mining of the relationships between historical sales and events. This estimates the traffic lift during promotional or holiday events. The estimation of traffic lifts occurs at different Escalation levels which are aggregated sales at different merchandise and location hierarchy levels.

3. Elasticity Estimation is generated from calculations of historical demand sensitivity based on markdown or promotional prices changes. The estimation of elasticity occurs at different Escalation levels which are aggregated sales at different merchandise and location hierarchy levels.

4. Seasonality Estimation utilizes various machine learning techniques to understand the time series in which data experiences regular and predictable changes that recur every calendar year. This reflects a sales pattern that repeats based on the time of year (season). The estimation of seasonality occurs at different Escalation levels which are aggregated sales at different merchandise and location hierarchy levels.

5. Base Demand Estimation calculates the average demand signal after the filtering of data, accounting for traffic lifts, and removal of elasticity and seasonality effects. This is generated at the same merchandise and location hierarchy level as the required forecast.

6. Forecast Generation calculates the demand forecast combining the output of all previous steps. This forecast multiplies base demand, seasonality, elasticity, and traffic lifts. In this manner, changes in seasonality, elasticity, and events can be evaluated and combined with the de-seasonalized base demand.

Hierarchical Bayesian Method

Hierarchical Bayesian method is useful for retailers wanting to optimize regular price of items. This is a Bayesian-based approach that uses regular price as one of the features to generate forecast. Typically, the data requirement for this method is more than one year of sales and prices.

The stages for Hierarchical Bayesian method:

1. Data Preparation defines the levels at which data is aggregated on merchandise, location/price zone, customer segment, and time dimensions and the duration of the data used for parameter estimation.

2. Preprocessing filters the historical data and makes the first determination of item eligibility.

3. Training enables the user to select the configurations and features that can be used for model training. Note that some features are mandatory.

4. Forecast Generation generates the ML models that can be later used by applications consuming demand forecasts.