Before you train a model, beat a spreadsheet
A machine learning project should start with the dumbest method that could work: last month's number, a simple average, a rule someone already uses. If the model cannot beat that, it should not ship.
Almost every business that asks for machine learning already has a way of making the same decision. The store manager orders roughly what sold last month. The credit team approves anyone with a salary above a threshold. The planner adds ten percent before Eid. These rules are not glamorous, but they work well enough that the business is still running.
That existing rule is the bar a model has to clear. Not zero, not a random guess: the thing people already do. We write it down and measure it before anyone trains anything.
What a baseline is
A baseline is the simplest prediction that makes sense for the problem. It costs nothing to build and it is easy to explain, which is exactly why it is useful.
- Forecasting sales: next week equals the same week last year, or the average of the last four weeks.
- Predicting churn: customers who have not ordered in 60 days will leave.
- Classifying support tickets: a list of keywords mapped to departments.
- Pricing: the median price of similar items listed this month.
Why it saves projects
Measuring the baseline first changes the conversation. Instead of "the model is 87 percent accurate", which means little on its own, the result becomes "the model cuts forecasting error from 22 percent to 14 percent compared with how you plan today". That is a number a manager can turn into money.
It also protects against a common failure: a model that looks impressive but only matches what a simple rule already does. When the gap is small, the honest recommendation is often to improve the rule, not to deploy a model that someone has to maintain.
If the trained model is only a little better than a simple average, the average may be the better product.Measure it the way it will be used
Testing it fairly
A baseline is only fair if both it and the model are tested the same way. The test must look like real use, or the numbers will flatter the model.
- Hold out time, not random rows: train on January to September, test on October to December. The future is what the model will face.
- Hold out whole customers or stores when the model will meet new ones.
- Use the measure the business cares about: money lost on stock-outs, not just average error.
- Look at the worst cases, not only the average. A forecast that is usually right but misses Eid completely is not usable.
Common problems, a sensible first baseline, and what a model has to add to be worth it.| Problem | First baseline | What the model must add |
|---|
| Weekly demand | Same week last year | Promotions, holidays, new items |
| Customer churn | No order in 60 days | Earlier warning, fewer false alarms |
| Fraud flags | Amount above a threshold | Catching small, unusual patterns |
| Lead scoring | Source of the lead | Ranking within each source |
When the honest answer is "keep the rule"
Sometimes the data simply does not hold enough signal, and no model beats the simple rule by a useful margin. That is a real result, not a failed project. It saves the business from maintaining a model that adds nothing, and it often points to the data that should be collected next.
A model is a cost. It has to earn its place against the rule it replaces.
Frequently asked questions
What if we do not have an existing rule?
Then we build the simplest sensible one together, usually in an afternoon, and measure it. It becomes the bar the model has to beat.
How much better does a model need to be?
Enough that the gain pays for building and maintaining it. For a forecast that drives ordering, a few points of error can be worth a lot; for a report nobody acts on, nothing is worth it.
Can the baseline and the model run side by side?
Yes, and that is how we prefer to launch. Both make predictions for a few weeks, and the team sees which one was closer before switching.
What data do you need to measure a baseline?
Usually the same history the model would learn from: past sales, past decisions, past outcomes. A year or two of records is enough for most forecasting and scoring problems.