Stock Predictor.
An applied-ML POC for figuring out whether modern models can find usable edges in public market data — and what a real workflow on top of that edge would actually look like.
Most "prediction" tools are demos in a dress.
Markets are noisy by design. Most ML pipelines that claim to predict prices either (a) overfit on a thin slice of history, (b) leak future information into training, or (c) work on backtests and quietly fail the moment they meet a live data feed. The interesting question isn't "can a model predict the next tick" — it's "can a model produce a signal that, after fees and latency and slippage, is still worth acting on."
That's what this POC is for. Not to ship a robo-advisor, not to compete with quant funds. To find out, on a small budget and with public data, whether the answer is yes — and if it is, what the smallest real workflow around that signal would look like for someone managing their own portfolio.
In plain English.
A research pipeline that takes a basket of equities, pulls public price and fundamental data, builds features at multiple timescales, trains a family of models, and reports honest out-of-sample performance — net of trading frictions. Then, on top of whatever survives that gauntlet, a small workflow that turns the signal into something a human can act on, with clear stop conditions and a fully auditable trade history.
Important: this is research, not financial advice. If the results don't survive, that's a result too — and an honest one.
Three possible landing zones.
Option A — the signal works, but only marginally: writeup, open-source the pipeline, publish the methodology so others can build on it.
Option B — the signal works and is meaningfully durable: a small tool I run for myself, with the option to share access carefully.
Option C — the signal doesn't survive proper testing: an honest writeup of why, and what the field gets wrong about backtests. Often the most useful kind of result.
Where we are today.
Data pipeline scaffolded. First batch of features built. Baseline models trained. Honest out-of-sample evaluation in progress — including the friction-aware kind that most published backtests skip. No production trading. No customers. Updates will land on this page as the experiment progresses.