Twenty years ago, we were sure to tackle forecasting and all the potential failures that could arise from the selection of the wrong method or the misapplication thereof in one of our first forecasting posts where we contrasted the two primary techniques in use at the time: judgemental forecasts and statistical forecasts.
In summary, we noted that:
| Forecast Type | Strengths | Weaknesses |
| Judgemental | * Environmentally Responsive * “Inside” Information * Responsive to “One-time” Events |
* Cognitive Limitations * Biases |
| Statistical | * Objective * Consistent * No Data Limits * Considers All Variables |
* No Real Time Reaction * Dependent on Model Formulation * Not Friendly to “Soft” Information * Technical Understanding Required |
The methodology that was typically used in practice was that a statistical model was run and then an expert adjusted it to their liking. Now, this usually worked okay if the organization had a real expert and access to up-to-date soft data, but otherwise, it was a shot-in-the-dark and poor forecasters would saddle the organization with stock-outs (and the ever-present 8% stockout rate) or excess inventory that would have to be fire-sold in the best case, and simply disposed of in the worst case. (Which is very costly from multiple angles, especially if you have to bury 700,000 cartridges in a landfill, which is where Atari’s E.T. found home in 1983.)
Things have changed in the last two decades. Now the average forecaster has two more options, which we really need to discuss: AI and Blended.
| Forecast Type | Strengths | Weaknesses |
| AI | * flexible model selection * flexible solution methodology * flexible interpretation of results |
* random model selection * performance limitations * hallucinations |
| Blended | * dependable models * multiple scenarios * balanced solutions that minimize downsides |
* computation time * data requirements * scenario weighting |
The AI option in particular is really concerning. Given that Gen-AI is, at best, artificial idiocy, you have no idea if the model it selects is appropriate, and even if the model it selects is reasonable, AI has never outperformed the classical statistical model equivalent, and then when you ask it for final interpretation, it can sometimes get the interpretation completely wrong. Generally speaking, modern (Gen) AI worsens forecasts.
In addition, the fact that most S&OP platforms still aren’t using balanced blended models is equally concerning. Considering that taking multiple models into account to produce a blended forecast reduces the chance of (significant) stockout and overstock when the reality doesn’t match the most likely scenario, blended should be the first choice. Especially since it’s better to be a few units short than have 10,000 extra or a few months of extra inventory that will eventually sell then be without for three months (due to long lead times).
In fact, the best method is blended models guided by human judgement that can identify the most likely scenarios, validate the best models, and then let the raw computational power of todays platform compute a solution that is near optimal across all of the different forecasts, rather than over optimizing for the wrong forecast. When you don’t know what the actual demand will be, it’s better to go with a blended solution that might be 2% off of optimal than pick the wrong scenario and end up with an “optimized” forecast that is 20% off.
Despite all of our advances, forecasting is still as much art as science, and only expert human judgement on blended models will produce sound forecasts in today’s rocky economic environment.





