Unlocking the Forecast Universe
Supply Optimization is a well-established concept in supply chain management, encompassing decisions related to procurement, manufacturing, storage, and shipping to maximize revenue or profit or minimizing costs. Forecast Optimization, although equally impactful, is less widely recognized. This article aims to clarify what Forecast Optimization is and how it operates, emphasizing its ability to autonomously minimize forecast errors, bias, and variability, thus substantially reducing the cost associated with forecast errors.
Defining the Forecast Universe
To understand Forecast Optimization, we need to explore the "Forecast Universe," which encompasses forecasts from several sources:
Machine Forecasts: Traditional Models include Exponential Smoothing, ARIMA (AutoRegressive Integrated Moving Average), Croston's method, and others that rely on historical data to predict future values. Machine Learning Models utilize advanced techniques such as Random Forest, AdaBoost, and Neural Networks that can capture complex patterns and relationships in data. Driver-Based Models forecast based on specific drivers or factors that influence demand, such as market conditions or promotional activities.
Human Forecasts: A major drawback in using Statistical Forecasting as a baseline for a fresh cycle is the complete disregard for human overrides made in the prior cycle. Humans are less likely to work from Statistical Forecast as a baseline to which they have to continually re-input their last cycle overrides all else remaining the same. vyan.ai considers Prior Cycle Human Forecasts, such as Customer Forecasts, Sales Forecasts, Marketing Forecasts, Demand Planner Forecasts, Consensus Forecasts, etc., while generating a baseline signal for the next cycle.
Hierarchy Levels: Businesses usually organize their forecast data across various planning hierarchies allowing forecasts to be reviewed at different levels of aggregation. Product Hierarchy may consist of Product ID, Product Group, Product Family, and Business Unit. Customer Hierarchy may consist of Customer ID, Customer Group, and Customer Region, facilitating forecasts at various customer-related levels. Forecasts can be generated at granular levels (e.g., Product-Customer) or aggregated levels (e.g., Product Group-Customer Group). Aggregated forecasts are often less noisy and more stable, making them useful for higher-level planning. Such aggregated forecasts are then broken back down to the granular level (Product-Customer ID) using system generated split factors.
Time Buckets: Forecasts can be made in different time intervals, such as weekly, monthly, or quarterly. More aggregated time buckets (e.g., monthly or quarterly) generally offer greater stability compared to more granular time periods (e.g., weekly), which can be more volatile.
Lags: Forecast combination must be not only time series specific (as different SKU/Customers will have different forecastability across humans and machines), but also lag specific. AI Optimal Forecast retains all the value-adding overrides made by each function at specific lags (sales may have better insights for the current quarter, marketing may have better insights further out).
Candidate Forecasts in the Forecast Universe
Combining various dimensions—human forecasts, machine forecast models, hierarchy levels, and time buckets—creates a diverse set of candidate forecasts within the Forecast Universe. For instance, to generate forecasts at a Product-Customer-Week level, we might also evaluate forecasts at broader levels such as Product Group-Customer Group-Month and Product Group-Customer Group-Quarter, using different models like ARIMA and AdaBoost.
Ensemble Methods
To enhance forecast accuracy, we use ensemble methods that combine different forecasts:
Full Ensemble: Averages all candidate forecasts, assigning equal weight to each regardless of their historical performance.
Top Ensemble: Selects the top 50% of forecasts based on their performance metrics (error, bias, variability) over the entire training period.
Best Training Forecast: Chooses the single forecast with the best performance metrics across the entire training period.
Recent Ensemble: Picks the top 50% of forecasts based on their performance over the most recent historical cycles.
Random Ensemble: Randomly selects 50% of the forecasts without considering their historical performance.
Hybrid Ensemble: Combines forecasts from Full Ensemble, Top Ensemble, Best Training, and Recent Ensemble methods to leverage their strengths.
Hyperparameters
AI Optimal Forecasting occurs on top of all the candidate and ensemble forecasts using a meta-learning model. Key hyperparameters in Forecast Optimization include:
Train vs. Test Split: The division of historical data into training and testing sets to evaluate forecast performance.
Overall Score (Loss Function): Weights assigned to different metrics like error, bias, and variability to assess forecast quality.
Meta-Learning Approaches: Techniques such as Linear Regression, Random Forest, and XGBoost, which determine how forecasts are generated and optimized.
How AI Optimal Forecasting Works
The vyan.ai forecasting engine integrates all candidate forecasts to produce a time series-specific and lag-specific Optimal Forecast. This system enables demand planners to review detailed contributions from both human and machine forecasts, providing insights into their respective impacts on the overall forecast.
The engine optimizes forecasts according to a configurable Objective Function, which balances accuracy and stability metrics to minimize forecast error costs. For instance, it dynamically adjusts the influence of a Prior Cycle Sales Forecast based on its effectiveness across different forecast lags, scaling down its impact as the forecast horizon extends from the short term to the mid-term. It ignores delta changes from Prior Cycle Consensus Forecast if such changes are within the noise level for a given time series. This is the same concept as discouraging frivolous insignificant overrides to improve signal stability across cycles, which helps reduce 'MRP Nervousness', a term used to signify large variability across cycles leading up to a future month.
Benefits
AI Optimal Forecast provides a starting point in your forecast cycle, which has the most potential to be adopted. Past cycle human overrides are retained (when such overrides have usually worked), the most accurate signal is made available across all models / levels / time buckets, and errors / bias / variability are all considered in an attempt to reduce the cost of forecast error and unnecessary churn across cycles. Such an Optimal Forecast can remove most of the avoidable error leading to a significant improvement in cost to serve and reduction in lost sales.
Learn More
We invite you to join our upcoming webinar on AI Optimal Forecasting, designed for Demand Planners, Analysts, and Data Scientists focused on achieving accurate and unbiased forecasts across cycles. Register here.
