Beat Nothing First: The Baseline Test Every Trading Model Must Pass

Every trading model should first be asked the rudest possible question: are you better than nothing?

Not better than the market. Not better than a hedge fund. Better than the dumbest rule a child could write down. If the answer is no, the model is done — however sophisticated its architecture, however glowing its backtest.

The two baselines that run on a bare CSV

You hand over a file of predictions. No price history, no extra data. Two honest baselines can still be computed from that file alone:

Always-bullish. Predict “up” every single time. Its score is simply the share of up-moves in the sample. If up-moves are 55% of your data, always-bullish scores 55% — with zero intelligence, zero parameters, and zero effort.

No-change. Predict that tomorrow looks like today. It captures persistence in the data — the tendency of markets to keep doing what they just did.

If your model’s accuracy cannot beat both of these, it has added nothing. A model claiming “55% accuracy” on a sample where always-bullish scores 55% is not a model. It is a complicated way of saying “up.”

Why this test is skipped

Because it is embarrassing. Sophisticated models with impressive backtests fail it more often than anyone admits — the model’s real skill is frequently just riding the base rate of up-moves, dressed in mathematics. Vendors skip the baseline because the baseline is the one comparison they cannot spin.

And note the boundary honestly: on a bare CSV, only these two baselines run. The other baselines people talk about — the ones that need price history — cannot be computed from predictions alone, and any report that claims otherwise is faking them. A bare CSV runs 2 baselines. Five need price history. That sentence is a lie detector for vendor reports.

How to run the test yourself

Take the model’s predictions and compute its accuracy from the raw predictions — never from a claimed-accuracy column in the file, which is trivially easy to inflate.

Compute always-bullish and no-change on the same data.

Subtract. The model’s edge is the margin above the best baseline, not the raw accuracy number.

If the margin is small, ask whether it survives costs, walk-forward folds, and a leakage gate. A 2-point margin over always-bullish rarely survives contact with reality.

BuildIT Index runs this test first, mechanically, on every bare CSV: recompute from raw predictions, score the two baselines, and only then talk about verdicts. In its verified demo, the sample model’s 69.33% recomputed accuracy is reported against those baselines — because a number without a baseline is a story, and stories are what backtests sell.

Beat nothing first. It is the cheapest test in trading, and it kills the most expensive mistakes.