Touchless Forecasting with Machine Learning within and outside SAP IBP

This blog post focuses on Touchless Forecasting: an enabler of demand forecasting optimization to produce the most optimal and granular forecasts by SKU/Location by month autonomously. This removes the process inefficiency and human bias challenges with most demand planning processes today.

Touchless Forecasting approach with SAP IBP in a nutshell:

• Develop the best forecast within SAP IBP across Traditional Forecasting (Exponential Smoothing, ARIMA, Croston, etc.) and Machine Learning approaches (Gradient Boosting of Decision Trees).

• Ingest forecasts from external Machine Learning / Deep Learning Models into SAP IBP (CNN, CatBoost, Prophet, Ridge Regressor, Random Forest Regressor, etc.)

• Optimally and autonomously blend all the statistical forecasts (Traditional or ML) and human forecasts (Sales Forecast, Marketing Forecast, Consensus Demand Plan, etc.) to produce the most accurate Touchless Forecast in SAP IBP by SKU/Location by Month.

Typical Business Challenges with Demand Planning in the Semi-conductor Industry:

• You must have heard about severe shortages of semi-conductor chips. Bookings (customer requested dates and quantities) are significantly disconnected from Shipments (confirmed dates and quantities; or actual shipments).

• Revenue forecasting process comes into focus in such a supply constrained environment, as booked demand largely loses meaning: it's all about the ability to service a subset of this booked demand, and yield revenue from it.

• Multiple semi-conductor finished goods can be delivered from the same underlying die part, this further complicates decision making to allocate constrained wafer/die part supply to competing finished goods and across customer demands.

• Customer Hubs (VMI Hubs) present an additional challenge to planning demand, as there are no firm bookings to taken into account when finalizing demand plans. Customer hub pulls are reflected as back to back sales order entry and shipment transactions.

• Data quality issues with Channel Inventory and Sales Out data (what Channel sells to end customers): Erratic frequencies of incoming data loads may result into 3 weeks of Sales Out data received along with current channel inventory (missing snapshots of channel inventory at the end of every week).

• Circular Planning: Demand Planners need to bake in overall supply constraints as they perform customer allocations. However, such allocations to specific customers / finished goods then need to be taken through the supply planning pass again to ensure all the components in the Bill of Material are available to support this allocation.

• Revenue Management: Top down revenue plan overrides need to be disaggregated to SKU level without violating overall supply constraints.

• Product transitions fragment demand history and make it harder to develop a good quality statistical forecast. Need to apply reference product history not just by SKU, but by SKU/segment, as new products are launched across multiple markets or customer segments. New product phase-in curves can seemingly last forever, as some customers take a long time to switch to the next-generation products. Launches are quite volatile, as OEMs may push the launch of their new products due to other component shortages. Supply constraints have to constrain demand forecast per historically observed new product - old product mix percentages.

• ASP impacts on Unit Forecasts: Average Selling Price ASP changes can impact demand non-linearly. Pricing policies and their non-linear impacts on forecasted demand have to be baked into the statistical forecast. Multiplying unit forecast with forecasted ASP is not enough, we need to model the impact of forecasted ASP on the unit forecast itself.

• Pandemic has caused extreme disruptions due to both demand and supply shocks. Statistical Forecasting with traditional extrapolatory approaches (Exponential Smoothing, ARIMA, etc.) is relatively ineffective. Machine Learning models are needed with the right demand drivers to capture the recessionary demand environment, constrained supply environment, etc. Machine Learning is harder to scale given its heavy focus on feature engineering, experimentation, and hyperparameter finetuning needed to deliver a high-fidelity signal that planners can understand and trust.

• Machine Learning is not the silver bullet in all cases. Traditional forecasting may deliver better results for some SKUs in some horizons, while human judged forecasts may be superior for other SKUs in other horizons. Given the plethora of Machine Learning Models. Traditional forecast models, and human forecasts, it is a real challenge for Demand Planners to define SKU level demand plans across the next four quarters across all SKUs within the 3-5 business days of time available in the monthly planning cycle to develop a demand forecast. Focus automatically goes to just one quarter out and just a few high-value SKUs, new products, and problem SKUs. This manual approach is neither scalable nor effective. Upstream contract manufacturers and component suppliers need a high-fidelity 4-quarter view of demand to plan their supply chains, else upstream supply chain constraints become a self-fulfilling prophecy.

Touchless Forecasting Solution Approach:

• Develop the best possible forecast across the entire forecast horizon (say 4-6 quarters). Develop such a high-quality forecast in a granular fashion: by SKU or by SKU/location.

• Choose the best signal by SKU by Quarter from a long list of competing forecasts: traditional forecast models, Machine Learning forecast models, and human forecasts. Have an autonomous unbiased approach to forecast selection: the forecast with the least error in the last 2 quarters wins. This selection has to be lag specific: we may choose to default the sales forecast for some SKUs in the next quarter, but may choose to use Gradient Boosting ML forecast 3 quarters out for the same SKU.

• The selection is based on the machine or human forecast with the highest Forecast Value Add (FVA) in the corresponding time frame (1 quarter out, 2 quarters out, 3 quarters out, etc.). The FVA calculation repeats every month to ensure the most dynamic view of FVA Winners is baked into the overall Touchless Forecast.

• You can visit GitaCloud YouTube Channel to review similar innovations and IBP explainer video content.