Causality: Forecasts Measured in Profit, Not Percentages

Causality is the software implementation of the Predictive Value Framework™ (PVF), a proprietary methodology developed by Dr Peter Catt, one of New Zealand’s most published practitioners in sales forecasting.

​

Every forecasting problem can be defined by three critical attributes. Getting these right determines whether machine learning creates value or quietly burns budget. PVF is designed to prevent the waste you often see in generic AI roll-outs, by forcing discipline around what can be predicted, what it's worth, and what can be influenced.

​

1) Forecastability: is there a signal?

Before we invest time in forecasting, we determine whether your data is genuinely predictable. We audit each time series up front using evidence-based indicators, including entropy measures, seasonality, and trend.

 

Machine learning is powerful, but it cannot reliably predict behaviour that is effectively random. When a series is highly irregular and lacks consistent structure, models tend to "see" patterns that do not hold up in the real world. In those cases, the sensible response is usually operational, for example, using buffers, safety stock, or service-level policy, rather than spending budget on a solution that can't add value.

 

We quantify this predictability using entropy and mutual information, which measure how much exploitable structure exists in the series at each forecast horizon (Catt, 2009, 2014, 2026).

​

2) Cost of Forecast Error: what is at stake?

Traditional forecasting focusses on generic accuracy scores that treat every error as equal. But improved accuracy does not automatically translate to better business outcomes, the relationship is far more complex. In practice, the cost of being wrong is rarely symmetrical: over-forecasting (excess stock, holding costs, obsolescence) carries different penalties than under-forecasting (lost sales, stock-outs, service failure).

We quantify that financial asymmetry in your context using the Cost of Forecast Error (CFE).

 

CFE operates as a governance loop around ML investment:

​

• Before modelling, we calculate the baseline CFE to identify where forecast improvement will generate financial returns, preventing investment in models for products where the cost of error is trivial, regardless of achievable accuracy gains.

• After deployment, CFE measures the value delivered in the same financial terms used to justify the investment, accountability in your organisation's own language.

 

Service level targets feed directly into CFE calculations, but they must be set independently of the forecasting process. Research shows that forecasters unconsciously anchor predictions toward service targets, systematically biasing demand forecasts. PVF addresses this by separating service level governance from forecast generation so that strategic targets influence inventory policy through CFE, not by contaminating the underlying predictions. The result is forecasting governed by business value, not statistical metrics (Catt, 2007, 2008a, 2008b, 2010).

​

3) Causal Machine Learning: can we intervene?

Most forecasting tools are passive: they extrapolate history and stop there. Causality goes further by applying machine learning techniques grounded in formal causal inference theory to understand why demand moves and whether you can change the outcome.

 

By separating correlation from causation, we identify controllable drivers such as pricing, promotions, supply constraints, and operational changes. That turns forecasting into a decision tool: instead of merely accepting a forecast, you can test the likely impact of specific actions and choose the best commercial response (Catt, 2017).

 

Where causal levers exist, CFE-governed investment in Causal ML is justified. Where they do not the emphasis shifts to CFE-optimised buffers rather than intervention strategies.

  

​

Forecasting for Profit

The Predictive Value Framework™ shows you how to measure and manage what actually matters, the financial cost of being wrong.

​