When Spreadsheets Beat Machine Learning

An unfashionable position

We build machine learning systems for a living, and a fair number of the enquiries we get would be better served by a spreadsheet. Not a lazy one. A well-structured workbook with clear inputs, sensible formulas and someone who owns it.

That is not modesty. A model that costs 30,000 pounds to build and a few thousand a year to maintain has to beat the spreadsheet by enough to pay for itself. Surprisingly often, it does not.

Five situations where the spreadsheet wins

1. There is not much data. A business with 40 products and two years of monthly sales has 960 data points. A seasonal average with a manual adjustment will match most models on that, and everyone can see how it works.
2. The pattern is already known. If your sales manager can explain that demand rises 30% in November and dips after Easter, write that down as a formula. A model would spend a lot of money rediscovering it.
3. Decisions are infrequent. A quarterly pricing review or an annual budget does not need automation. It needs good thinking with the numbers in front of it.
4. People must adjust the logic. When a planner needs to say 'we have a promotion that week, add 15%', a transparent sheet is easier to override than a model.
5. The world just changed. After a big shift in your market, historical data is misleading. Human judgement in a spreadsheet adapts faster than a model trained on the old world.

A worked example

Take an illustrative specialist food wholesaler forecasting orders for 120 products. Their current spreadsheet uses a 12-month moving average, a seasonal index per product family and a column for the buyer's manual adjustment.

A machine learning model trained on their history might reduce forecast error by a few percentage points on the fast-moving lines and do no better on the slow ones. If better forecasts on those lines save a couple of thousand pounds a year in waste and stockouts, the model does not pay for itself for a decade. Improving the spreadsheet, perhaps adding a proper seasonal index and a check against last year, would capture much of that gain in a week.

Change one number, 120 products to 12,000 across six warehouses, and the answer flips completely. Nobody can maintain seasonal indices for 12,000 lines by hand. That is where a model belongs.

Where machine learning pulls ahead

The pattern is volume multiplied by complexity. A small accuracy gain applied to thousands of decisions a day is worth a model. The same gain applied to twelve decisions a year is not.

The real problem is often not the method

When a business says its spreadsheet is not good enough, the issue is frequently something other than the forecasting method. Nobody trusts the input data. Three versions are emailed around. A formula was broken in March and nobody noticed. The person who built it left.

Those are real problems, but machine learning solves none of them. A model fed from the same untrusted data will produce untrusted predictions with more decimal places. The fix is often moving the data into a proper system with a single source of truth, which we cover in signs your business has outgrown spreadsheets.

Making the spreadsheet dependable

If the spreadsheet is the right tool, it is worth making it a good one.

• Separate inputs, calculations and outputs onto different sheets
• Protect formula cells so they cannot be overwritten by accident
• Keep one master copy in a shared location, with version history switched on
• Add a simple check row that compares this month's total against last year's, so broken formulas show up
• Record forecast against actual every month, so you know how good the method really is
• Write a short note explaining the logic, for whoever inherits it

That last point about recording actuals has a bonus. In two years you will have exactly the evidence needed to judge whether a model would do better, which is the only honest way to decide.

When we recommend moving on

The signals are fairly consistent: the number of items or decisions has grown past what people can review, errors are costing noticeably more than a build would, and the same adjustments are being made by hand every week, which means there is a pattern a model could learn. A related question is when a simple rule beats a model, which applies just as much inside a spreadsheet as in software.

When SpiderHunts is asked about machine learning for forecasting or scoring, we start by comparing the model against the existing spreadsheet on recent data. If the spreadsheet holds its own, we say so, and usually suggest the three or four changes that would make it more reliable.

Frequently asked questions