Demand Forecasting Without a Data Team

Every mid-market business forecasts demand. Very few forecast it well.

The typical setup: one spreadsheet, one owner, last year's numbers plus a growth percentage, adjusted by gut feel for known events. It works until it does not, and when it fails it fails expensively: cash tied up in the wrong stock, expedited freight to cover a stockout, or a shift roster built for demand that never arrived.

The standard advice has been to hire a data scientist. For a business turning over 5 to 100 million pounds, that advice has always been impractical. What changed in the last eighteen months is that it is also no longer necessary.

The spreadsheet ceiling

Spreadsheet forecasting is not bad because spreadsheets are bad. It is bad because of four structural limits that no amount of skill inside the file can fix.

It extrapolates one variable. Last year's sales, uplifted. Real demand responds to price, promotions, weather, competitor activity, and calendar effects at the same time. A single trend line cannot separate them, so every driver collapses into one number the forecaster then overrides manually.

It handles seasonality badly. Most spreadsheet models apply a flat monthly index. Real seasonality is multi-layered: day of week, week of month, public holidays, school terms, and an annual cycle that shifts by a few days each year. Averaging those together produces a forecast that is wrong in a predictable direction almost every week.

It produces a single number with no confidence interval. A forecast of 4,200 units tells you nothing about risk. A forecast of 4,200 units with an 80 percent range of 3,600 to 4,900 tells you exactly how much safety stock the month requires. Planning teams that only receive point estimates end up setting buffers by policy rather than by evidence, which is how businesses simultaneously hold too much of the wrong stock and too little of the right stock.

It has a single point of failure. The person who understands the overrides. When they are on leave, the forecast degrades. When they leave the business, the logic leaves with them.

None of this is a criticism of the people running those files. They are usually the most commercially astute person in the building. The problem is that they are being asked to do statistics by hand, weekly, alongside their actual job.

What a forecast error actually costs

Forecast accuracy is usually discussed as a percentage, which makes it feel abstract. Convert it to money and it stops being abstract quickly.

Take a distributor with 20 million pounds of annual revenue, a 25 percent gross margin, and 4 million pounds of average inventory. Assume a current forecast error of 30 percent, which is typical for spreadsheet planning on a multi-SKU catalogue.

Overstock cost. Excess inventory driven by forecast error ties up working capital and generates holding cost at roughly 20 to 25 percent of inventory value per year once you include capital, storage, insurance, shrinkage, and obsolescence. If a fifth of that 4 million pound inventory position exists purely as a buffer against forecast uncertainty, that is 800,000 pounds of stock costing around 180,000 pounds a year to hold.

Stockout cost. The other side of the same error. A 2 percent stockout rate on 20 million pounds of revenue is 400,000 pounds of demand that either arrives late or goes to a competitor. Even attributing only half of it as genuinely lost, at 25 percent margin, that is 50,000 pounds of gross profit.

Expedite cost. Every emergency replenishment carries a freight premium and internal handling cost that never appears in any forecasting business case because it is buried in logistics.

Cutting forecast error from 30 percent to 18 percent does not eliminate those numbers. It reduces them roughly in proportion. On the illustrative business above, that is a six-figure annual swing from a change that touches no headcount and no warehouse.

The assumptions above are illustrative, not a client result. The point is the method: before evaluating any forecasting tool, work out what one point of accuracy is worth in your business. If the answer is small, do not buy anything. If the answer is large, you now have a budget and a benchmark.

Why you no longer need a data team

The old model of demand forecasting required a model per series. You had 3,000 SKUs, so you trained 3,000 models, tuned them, monitored them for drift, and retrained them on a schedule. That is a genuine data science and MLOps workload, and it is why forecasting stayed the preserve of businesses large enough to fund a team.

The shift is that forecasting now has pretrained foundation models, in the same way language does. These models are trained once on enormous and diverse collections of time series, and then applied to a series they have never seen before, producing a forecast without any training run on your data. In the field this is called zero-shot forecasting.

Practically, that removes four things from the project:

No training pipeline. You are not fitting a model, so there is nothing to schedule, monitor, or retrain when it drifts.

No feature engineering. Classical approaches need you to construct lag features, holiday flags, and seasonality terms by hand. A pretrained model has learned those patterns from millions of other series.

No per-SKU maintenance. The same model serves the whole catalogue, including new SKUs with almost no history, which is precisely where classical methods perform worst.

No specialist headcount. The work becomes data plumbing and interpretation, which your existing analyst or ops lead can own.

What has not disappeared is the need for clean history. A pretrained model cannot invent signal that your data does not contain. If your SKU codes changed twice in three years and nobody mapped the old ones forward, that is still a project. It is a smaller project than hiring a team, but it is not nothing.

What good looks like

A short comparison of the three approaches most mid-market businesses choose between.

Spreadsheet. Setup time: already done. Data required: a sales export. Seasonality: crude monthly index. Confidence intervals: none. Maintained by: one person, manually, weekly. Best for: single-product businesses and very stable demand.

Classical statistical models. Setup time: weeks to months. Data required: two to three years of clean history per series. Seasonality: good, if configured per series. Confidence intervals: yes. Maintained by: an analyst who understands the model family. Best for: stable catalogues with long history and a team to look after them.

Foundation-model forecasting. Setup time: days. Data required: as little as several months of history, more is better. Seasonality: learned automatically, including multiple overlapping cycles. Confidence intervals: yes, as prediction quantiles. Maintained by: nobody, in the training sense. Best for: multi-SKU catalogues, new products, and teams without data science capacity.

None of these is universally correct. A business with forty products and flat demand should stay in the spreadsheet and spend the money elsewhere. A business with 3,000 SKUs, seasonal peaks, and a planner working weekends is leaving a lot on the table.

The 30-day validation

Do not buy a forecasting system on a demo. Demos run on the vendor's data. Run a backtest on yours. This takes four weeks and costs almost nothing.

Week one: assemble the history. Export at least 24 months of demand at the grain you actually plan on, usually SKU by location by week. Two rules matter more than anything else here. Use demand, not sales, so that periods where you were out of stock are marked as censored rather than recorded as zero demand. And freeze the SKU mapping, so a product that was renamed does not look like two products with half a history each.

Week two: establish your baseline. Take your existing forecast for the last six months and measure it honestly. Use WAPE, weighted absolute percentage error, rather than MAPE for catalogues with low-volume items, because MAPE distorts badly when actuals are near zero. Record the number. Most businesses have never measured this, and the exercise alone is often the most valuable part of the month.

Week three: run the challenger. Hold out the last three months, forecast them from the earlier data only, and score the result with the same metric on the same items. This is the only comparison that means anything. A model evaluated on data it has seen will always look excellent and will always disappoint in production.

Week four: convert accuracy into money. Take the error reduction and apply it to the cost model from earlier in this article. If a 12 point WAPE improvement translates into a six-figure inventory and service swing, the decision is straightforward. If it translates into 15,000 pounds, do not build it.

Two guardrails. Score by value band, not just overall, because a model that improves your slow movers and degrades your top twenty SKUs is a bad trade regardless of the headline number. And insist on prediction intervals rather than point forecasts, because the interval is what actually sets safety stock.

Where forecasting fails

Honest limits, because a guide that only lists upside is a brochure.

Sparse or broken history. Products with a handful of sales events, or catalogues where identifiers were not maintained, will not forecast well by any method. Fix the data first.

Genuine demand shocks. No model predicts a competitor's insolvency, a regulatory change, or a viral moment. What a good model does is recover quickly and tell you, through widening intervals, that it is now less certain. Treat a widening interval as a signal to intervene, not as a failure.

Forecasting things you cannot act on. If your supplier lead time is sixteen weeks and your forecast horizon is four, better accuracy changes nothing. Match the horizon to the decision it feeds. This is the single most common reason forecasting projects deliver no value despite working correctly.

Organisational override. A forecast that the commercial team does not trust will be overridden, and overrides usually make accuracy worse. Track override impact from day one. If manual adjustments consistently degrade the number, that is a conversation worth having with evidence rather than opinion.

Where to start

If you are running a multi-SKU catalogue on a spreadsheet and you have never measured your own forecast error, start there. It costs a week and it tells you whether any of this is worth pursuing.

If the number is worse than you expected, and it usually is, the next step is a backtest against a modern model on your own history before any commitment.

Agni Predict is our forecasting product for exactly this situation: pretrained forecasting applied to your history, with prediction intervals, no training pipeline, and no data science hire. You can see the deployment modes and join the waitlist on the Agni Predict page.

If you are not sure forecasting is your biggest constraint, run our free Business Diagnostic first. It analyses your real customer reviews against the problems you tell us about, and tells you where the actual bottleneck is. Sometimes it is forecasting. Often it is something cheaper to fix.

And if you want the wider commercial case for agentic and predictive AI in the mid-market, our Agentic AI ROI Playbook covers how we size and stage these engagements.