Baseline Models and Naive Benchmarks

A model predicting next-day volatility reports a mean absolute error of 2.1 volatility points. Is that good? Without a comparison point, that number means nothing at all — it could be a genuine breakthrough or it could be barely different from just guessing "the same as yesterday" every single day. Teams regularly spend months building an elaborate model, report an impressive-sounding accuracy figure, and never check whether a one-line naive rule would have scored almost as well. A baseline model is the deliberately simple, cheap-to-build predictor you're ethically obligated to check first, because it's the honest floor everything else has to clear.

An analogy: judging a shortcut without knowing the normal route

Imagine someone tells you they found a route across town that takes 22 minutes, and asks you to be impressed. You can't judge that claim at all until you know how long the obvious, no-thought-required route takes — if the direct route also takes 22 minutes, the "shortcut" saved nothing; if the direct route takes 45 minutes, it's a genuinely useful discovery. A baseline model is that obvious, no-thought-required route: it's not meant to be good, it's meant to be the honest reference point that tells you whether your fancy model's reported performance represents real skill or is barely distinguishable from doing the simplest possible thing.

Common baselines, by task type

For regression (predicting a number, like next-day volatility or a stock's return): the persistence baseline predicts tomorrow will equal today; the mean baseline predicts the historical average value every time, ignoring all features.

For classification (predicting a category, like whether a stock beats its sector): the majority-class baseline always predicts whichever class is most common in the training data; a stratified random baseline predicts each class with a probability matching its frequency in the training data, useful for sanity-checking metrics like AUC on imbalanced classes.

For trading strategies specifically: equal-weighting, a simple buy-and-hold on the relevant benchmark index, or a naive rule like "buy last month's winners" all serve as baselines a genuinely valuable model needs to beat after costs, not just in gross backtested return.

In plain English: a baseline uses almost no real modeling effort or insight, deliberately, so that any margin your actual model shows over it can be attributed to the modeling effort itself rather than to some property of the data or task that would have made even a trivial guess look decent.

Worked example: is 58% accuracy actually good?

A model predicting whether a stock beats its sector next week reports 58% accuracy on a held-out test set. The training data shows the stock beats its sector 55% of the time overall — so a majority-class baseline that always predicts "yes, it will beat the sector" would score 55% accuracy without using a single feature. The model's real, baseline-adjusted improvement is only $58\% - 55\% = 3$ percentage points, not the full 58% the raw number suggested at first glance. That's a much less impressive story, and it changes the conversation from "we built a good model" to "we built a model with a modest, real edge over guessing the more common outcome" — which is still potentially valuable, but requires a very different level of confidence before committing real capital than an unadorned 58% figure implied.

What this means in practice

Always compute and report the relevant baseline alongside any model's headline metric, in the same table, before anyone gets to react to the model's number in isolation. This is cheap — usually a few lines of code — and it protects against the single most common way research results mislead: a metric that sounds impressive purely because the underlying task's baseline happened to already be high (or low). It's also a useful early gate: if a complex model can't clear a simple, honest baseline by a meaningful margin, that's a strong signal to stop and reconsider the approach before investing further research time.