Machine Learning vs Statistics vs Spreadsheet Formulas

Three tools that get confused

A managing director asks for machine learning to forecast next quarter. Half an hour later it turns out the forecast lives in a workbook where someone copies last year's figures and adds 8%. What would help most is not a neural network. It is a properly built forecast with seasonality and a range, which a competent analyst can produce with ordinary statistics.

Spreadsheets, statistics and machine learning overlap, and vendors blur them further. Separating them saves money, because each costs a very different amount to build and maintain.

What each one is for

Spreadsheets: underrated until they are not

A well-built spreadsheet is a precise, transparent model of how you believe the business works. Pricing calculators, cash flow projections and commission rules are all perfectly suited to it. If you can state the logic, a formula implements it, and anybody can check it.

Spreadsheets start to hurt when many people edit them, when data is pasted in by hand every week, or when a formula hard-codes an assumption nobody revisits. Those are signs you need a database or proper software, not necessarily machine learning. Our article on signs you have outgrown spreadsheets covers that transition.

Statistics: when you want to understand and decide

Statistics is the right tool when the question is about the business as a whole rather than each individual case. It is built to tell you whether an effect is real or noise, and how sure you can be.

• Did the new checkout page actually increase conversion, or was it a good fortnight
• How much does a one-day delay in delivery reduce repeat orders
• What range should next quarter's sales fall in
• Is branch A's higher return rate a real problem or chance
• Which of these marketing channels is associated with better customers

The output is understanding: an estimate, a range and a level of confidence. That is exactly what boards and finance directors need, and it is often cheaper and more trustworthy than a predictive model. A lot of this work can be done with Python or R, and some of it in a spreadsheet by someone who knows what they are doing. Our note on Python data analysis gives a sense of what that looks like.

Machine learning: when you need a prediction per case

Machine learning earns its cost when you need a separate, good prediction for each of many cases, made repeatedly, and the patterns are too tangled to write down. Scoring every incoming lead. Estimating the delivery time for every order. Flagging every suspicious transaction as it happens.

It trades some transparency for predictive power and scale. Statistics might tell you that customers who log in less are more likely to leave. Machine learning combines login patterns with forty other signals to rank this morning's 4,000 accounts by risk. Both are useful and they serve different decisions.

Statistics asks what is going on and how sure we are. Machine learning asks what will happen to this one, and this one, and this one.

The line between them is blurrier than it looks

Linear and logistic regression appear in statistics textbooks and in machine learning courses. The maths is the same. The difference is in purpose and emphasis: statisticians care about whether each coefficient is meaningful and trustworthy, machine learning practitioners care mainly about predictive performance on new data.

For a business this is good news. A well-built statistical model can often be deployed as a predictive one, and a simple, interpretable model is frequently the best first version of a machine learning system anyway.

It also means the path between the three is gradual. A workbook that forecasts by adding 8% can become a statistical forecast with seasonality, which can later become a model that predicts demand per product per store. Each step should be justified by a decision that the previous step could not support. Skipping straight to the last one is how people end up paying for complexity they never use.

Choosing, in practice

1. If you can write the logic down and it rarely changes, use formulas or ordinary software
2. If you need to understand an effect, test a change or forecast totals with a range, use statistics
3. If you need individual predictions for many cases, repeatedly, with patterns too complex to write, consider machine learning
4. If you are not sure, start with statistics, because what you learn will shape any model you build later

At SpiderHunts a good number of enquiries labelled machine learning end up, after a short conversation, as a statistics or reporting job, or as custom software to replace a fragile workbook. We would rather sell you the right small thing than the wrong big one. When prediction really is the need, the machine learning service is where that work lives.

Frequently asked questions