Demand sensing won't fix a forecast nobody trusts

Every demand-sensing vendor pitches the same arithmetic: near-term forecast error down 30 to 50 percent, safety stock down, service level up. The technology is genuinely real — it uses point-of-sale data, open orders, weather, and web signals with machine learning to refine the zero-to-twelve-week forecast. The numbers are not wrong. The framing is.

The claims are easy to find. e2open states its demand sensing improves near-term forecast accuracy by 30 to 40% compared with traditional methods. Arkieva puts the reduction in short-term forecast error at up to 40 to 50%, with inventory accuracy up 15 to 20%. These ranges are credible in pilot conditions. They are also irrelevant to most enterprises' actual problem, because a demand-sensing improvement only matters if the resulting forecast is trusted and used.

Here is the data that should reframe the whole conversation.

In Steve Morlidge's study of more than 300,000 real-life forecasts across eight supply-chain companies, 52% were worse than a naïve random-walk forecast — worse, that is, than simply using last period's actual. Follow-up work across twenty-plus companies found 30 to 50% of low-level forecasts routinely worse than naïve.

Sit with that. If half your forecasts are already losing to "use last week's number," then a 30 to 40% demand-sensing improvement is recovering ground you should never have ceded. It is treating a symptom.

And the symptom has a cause that no algorithm addresses. In fourteen years I have seen exactly four root causes of forecast failure, none of them statistical:

• No single owner of the number. Sales, finance, marketing, and supply chain each keep their own forecast and litigate the differences in a meeting that resolves nothing.
• Incentives misaligned with accuracy. Sales is paid to hit quota, finance to hit budget, supply chain to hit service. None is paid to predict reality.
• Judgmental override that destroys value. This is the killer, and it is measurable.
• No feedback loop. Most companies never track which step — the engine, the analyst, the consensus meeting, the executive — degraded the forecast, so they cannot fix it.

On the third point, the research is brutal. A 2024 study in the International Journal of Forecasting found that judgmental adjustments improved accuracy for only just over half of stock-keeping units, and that positive adjustments were more likely to worsen performance. Demand sensing produces a better statistical forecast. If your organisation does not trust statistical forecasts, it will override the better one too — and the evidence says it will probably make it worse.

So before you buy demand sensing, run a Forecast Value Added analysis. Measure each step of your process against that naïve benchmark. If your current process is destroying value relative to a random walk — a near-coin-flip likelihood on Morlidge's evidence — the right investment is not a better algorithm. It is clear ownership of the number, accuracy-linked incentives, FVA-monitored overrides, and the discipline to remove the executive override that consistently destroys value.

Buy demand sensing after you have done that. Not before. A sharper forecast handed to an organisation that doesn't trust forecasts is an expensive way to generate a number people will still ignore.

---

Sources

• Morlidge, S. FVA and the Limits of Forecastability, IBF Amsterdam — cited in Gilliland, Tashman & Sglavo, Business Forecasting: Practical Problems and Solutions (Wiley, 2015).
• Fildes, R. et al. (2024). Forecast value added in demand planning. International Journal of Forecasting. sciencedirect.com
• e2open. Demand Sensing Software. e2open.com
• Arkieva. 6 Ways You Can Improve Forecast Accuracy with Demand Sensing. blog.arkieva.com
• Kinaxis. What is demand sensing? kinaxis.com