- AI
The baseline before the model: the rule that saves time and buys credibility
A moving average, a simple business rule, last month's value. Any model must beat that reference to ship — and the reference regularly wins.
· 4 min read · AzerOps
In a modelling project the first thing to build is not a model. It is the dumbest solution that addresses the problem, measured on the same metric and the same dataset as the model that comes afterwards.
What a baseline is
It must be trivial to explain and to implement:
- Demand forecasting: a three-month moving average, or the same month last year.
- Anomaly detection: a fixed threshold set by a business expert.
- Classification: the majority class, or an existing business rule already applied manually.
- Scoring: the spreadsheet score somebody already uses.
The common trait: nobody doubts what it does, and it could go to production in a day.
Why it changes everything
It gives the numbers meaning. 87% accuracy is neither good nor bad in itself. If the baseline scores 85%, your model brings two points at the price of permanent industrialisation and maintenance cost. If it scores 62%, the contribution is substantial. Without a baseline the discussion is impossible.
It protects against over-investment. The baseline regularly wins, particularly on short series, very noisy ones, or ones dominated by exceptional events. Finding that out in two days costs two days. Finding it out after three months of development costs three months.
It builds user trust. A planner who sees their provider declare that a model adds nothing on three product families out of seven accepts the model on the other four far more readily. That is counter-intuitive commercially and it holds up in practice.
It gives you a fallback. A model in production fails, drifts, or has to be disabled during an investigation. If the baseline is still implemented, switching back takes a minute. Otherwise you revert to a manual process nobody remembers how to run.
The rule to write into the contract
"No model is deployed to production unless it beats the baseline on the agreed metric, measured on a test set used in no training run."
One sentence. It changes the nature of the project: the provider no longer has an interest in delivering complexity, they have an interest in delivering a measurable result.
The usual objection
"If we set the bar too low, the provider will settle for the minimum." The opposite happens. The baseline is a floor, not a ceiling: it defines what is unacceptable, not what is targeted. What is targeted is the success threshold set at framing, which is a different number, usually well above.
What it means for your tender
Ask every provider you consult which baseline they propose and how they intend to measure it. Those with no prepared answer are telling you something useful about how they run their projects.