About Equity Doodle

Equity Doodle is an educational case study that turns stock-chart "completion" into a time-series forecasting problem. The segment you draw becomes a model's input context window, and the continuation it sketches is the forecast horizon. One global model is trained across ~530 assets — equities, ETFs, indices, futures, crypto, FX, commodities, rates — with daily history back to 1980 plus intraday panels, across many time resolutions, so it learns the general grammar of price paths rather than memorizing a single ticker.

How it works

• Your drawing is reduced to a fixed number of bars and converted to log-returns — so price scale and how many points you drew don't matter.
• The returns are z-scored (normalized to their own volatility) and fed to a global N-HiTS neural network with quantile regression — the winner of an honest bake-off against gradient boosting and TiDE on out-of-sample fan quality.
• The raw fan is conformally calibrated: quantile models are overconfident out of the box (the nominal 10–90% band covered only ~60% of realized moves), so the fan is widened by a factor fit on held-out data until the 80% band actually covers ~80%.
• The predicted return distribution is rescaled by the volatility you drew and integrated back into a price path. You see a jagged sample continuation (one plausible future), a faint median, and the shaded uncertainty band.

Because everything happens in standardized return space on an integer index, the system is time-scale invariant: draw at any resolution — intraday, daily, weekly — and any price scale, and it still works.

It runs entirely in your browser

The trained network is exported to ONNX (~1 MB) and executed locally with ONNX Runtime Web. Inference never touches a server: nothing you draw ever leaves the page and forecasts cost nothing to serve. (The one exception to "no API" is the live-audit page, which fetches market quotes through a tiny proxy — your drawings still stay in the browser.)

What the median and the bands are really telling you

A reliable chart pattern is, by definition, one whose shape dictates what comes next — the same setup producing the same outcome, again and again. If such time-scale-invariant technical patterns genuinely existed, this model would be ideally placed to expose them: trained across many assets and resolutions, it will have seen every common shape thousands of times. You would then notice two tell-tale signatures:

What you actually observe is the opposite. No matter how you wiggle the input, the median collapses toward a gentle drift, and the uncertainty band fans out wide. That is not a defect in the model — it's a direct, visual demonstration of how little of future price movement is determined by the shape of the recent past. The wide fan is the result: it's the efficient-market intuition, drawn. The honest backtests reflect the same story — directional accuracy at a coin flip on daily data (only modestly above on weekly), and point error no better than a random walk.

Why this matters as a study

The point was never to make money. It's to build the full pipeline end-to-end — data, a scale-invariant global model, an honest bake-off, calibrated uncertainty, and an interactive UI — and to let the model's own uncertainty show you why "the chart tells you where it's going" is mostly a story we tell ourselves.

Don't take our word for it: play Beat the Model. You get a real historical chart with the ticker hidden, you draw what happens next, and then the truth is revealed alongside the model's fan — with the score. Most people discover they can't reliably beat it. Neither can it reliably beat a coin flip. That's the lesson.