The reward of taking the next step to the summit lies in…
StAIrS&P 500 prediction and analysis powered by machine learning
What STAIR is
STAIR is a directional prediction platform for the S&P 500 via the SPY ETF. A daily pipeline ingests market data, technical indicators and macroeconomic variables to feed 72 models in production: nine algorithms across eight combinations of cadence (daily, weekly, monthly) and prediction horizon. All of them retrain automatically every month.
Each prediction can be checked against a backtesting engine that discounts real transaction costs, and against an investment simulator that reconstructs what would have happened with real money on the table: three investment strategies, entry and exit dates configurable from the year 2000, up to five periods with different contributions, initial capital, position sizing, transaction costs, short positions and your specific tax situation (IRPF, pyme or régimen general — Spain's personal income tax, reduced-rate small-business, and general corporate tax regimes, respectively). The results of each simulation can be explained in natural language by an AI assistant.
The honest result
And the result is that it doesn't work. Under nested walk-forward validation with temporal isolation, none of the 72 models robustly beat the baseline: simply predicting the historical mean outperforms almost everything else once transaction costs are discounted. This is the result one should expect under the weak-form market efficiency hypothesis applied to an ETF as liquid as SPY, and it is consistent with the literature warning that most positive backtests in financial machine learning are artifacts of overfitting.
STAIR publishes that result instead of hiding it. The goal was never to demonstrate abnormal returns: it was to build and document the infrastructure capable of measuring that honestly.
Context and contact
STAIR has run in production unattended since late June 2026: a daily pipeline, automatic monthly retraining and a public API. It is built and maintained by a single person. It started as a final degree project in Data Science and Engineering at the Universidade da Coruña.
Questions, comments, or found something that doesn't add up? Write to me at avarela@stair-tfg.me.
Technical blog
The blog publishes technical articles on the project's design and validation: architecture decisions, experiment results, and what didn't work.
Read the articlesWho built STAIR
STAIR was built by Alejandro Varela, working alone. The author's page covers the project's technical decisions and results.
About the authorWalk through STAIR from start to finish
Four sections, in the order a prediction is built: from the data that goes in to the money that would have been made or lost.