How to optimize demand forecasting outputs for business objectives

Demand forecasting is a critical process for businesses, particularly in retail, where accurate predictions of product sales can significantly impact operations. However, creating forecasts that truly serve business needs goes beyond simply predicting numbers. This is where metric weighting comes into play.

Why traditional demand forecasting falls short

Traditional approaches often fail to meet the complex needs of modern businesses. Let's dive into the key issues:

1. One-size-fits-all approach: Traditional forecasting methods typically aim for a single "best" forecast. However, businesses have multiple departments with diverse needs. A forecast that works well for inventory management might not be suitable for financial planning or staffing decisions. This one-size-fits-all approach often leaves some business needs unmet.

2. Intermittent demand challenges: Many products, especially in retail, don't sell every day. This intermittent demand poses a significant challenge for traditional forecasting methods. Models designed for smooth, continuous data often struggle with sporadic sales patterns, leading to inaccurate predictions.

3. Overemphasis on common metrics: Forecasting models are often optimized for common metrics like RMSE (Root Mean Square Error) or MAPE (Mean Absolute Percentage Error). While these metrics have their place, they can lead to counterintuitive results. For instance, with intermittent demand, a forecast of all zeros might score better on RMSE than a more nuanced prediction, despite being less useful in practice.

4. Lack of business context: Many forecasting methods focus solely on statistical accuracy without considering the broader business implications. They may produce forecasts that are mathematically sound but fail to capture important business dynamics or seasonality.

5. Difficulty in handling multiple time horizons: Businesses often need forecasts for different time horizons - daily for operational decisions, weekly for mid-term planning, and monthly or quarterly for strategic decisions. Traditional methods often struggle to provide accurate forecasts across these varying time scales.

6. Inability to balance conflicting priorities: Different departments may have conflicting needs from a forecast. For example, the sales team might prefer optimistic forecasts to set ambitious targets, while the inventory team needs conservative estimates to avoid overstocking. Traditional methods don't offer a way to balance these competing interests.

7. Lack of flexibility for changing business needs: Business priorities can shift quickly, but many forecasting systems are rigid and difficult to adjust. This lack of flexibility means forecasts can quickly become misaligned with current business objectives.

These challenges highlight why traditional demand forecasting methods often fall short in meeting the diverse and complex needs of modern businesses. They underscore the need for a more flexible, context-aware approach to forecasting that can align predictions with specific business objectives and handle the nuances of real-world demand patterns.

Introduction to metric weighting

Metric weighting offers a powerful solution to many of the problems faced in traditional demand forecasting. By allowing businesses to incorporate multiple evaluation criteria into their forecasting models, this approach provides a more flexible and business-aligned method of prediction. Here's how metric weighting addresses the key issues:

1. Multiple use cases: Most businesses have several applications for their demand forecasts, such as inventory management, staffing decisions, and improving customer experience. Each of these use cases may require a different focus in the forecast.

2. Balancing priorities: Metric weighting allows for the creation of a single, comprehensive forecast that balances multiple priorities. This prevents the confusion and potential conflicts that could arise from having separate forecasts for each use case.

3. Addressing specific challenges: Different metrics can address various forecasting challenges. For instance, in cases of intermittent demand (where products don't sell every day), traditional metrics like RMSE (Root Mean Square Error) may not provide the most useful forecasts.

4. Business value focus: By weighting metrics that align with specific business objectives, companies can ensure their forecasts optimize for the factors that drive the most value.

While metric weighting offers significant advantages, it's important to note that not all forecasting algorithms can effectively utilize this approach. Metric weighting requires a sophisticated optimization framework capable of handling multiple, potentially conflicting objectives simultaneously.

Implementation of metric weighting

To implement metric weighting:

1. Identify key business objectives and select metrics that align with each goal.

2. Assign weights to these metrics based on their relative importance to the business.

3. Use these weighted metrics to guide the selection or fine-tuning of forecasting models.

4. Regularly review and adjust weights as business priorities evolve.

See below as an example of some of the metrics businesses are able to adjust as they tune forecasts to their needs.

# Example weighting, untuned
metric_weighting = {
    'smape_weighting': 1,
    'rmse_weighting': 1,
    'spl_weighting': 1,
    'mage_weighting': 1,
    'mate_weighting': 1,
    'wasserstein_weighting': 1,
    'dwd_weighting': 1,
    'runtime_weighting': 0.05,
}

By adopting a metric weighting approach, businesses can create more nuanced, flexible, and business-aligned forecasts. This method transforms demand forecasting from a purely statistical exercise into a strategic tool that directly supports business objectives across multiple departments and time horizons.

Our friends at Nousot are currently the only forecast provider that support metric weighting. They've developed an end-to-end demand forecasting solution for their clients that can be run on Databricks or a traditional cloud environment. Learn more about their approach from this post:

Much of the content from this post was inspired by an original post from Colin Catlin, a co-developer of the demand forecasting solution by Nousot.