Dancesp/sma-zoomcamp-2026-data

SMA Zoomcamp 2026 project outputs

The public outputs of an end-to-end machine-learning trading project built for the DataTalksClub Stock Markets Analytics Zoomcamp 2026: hourly decisions on S&P 100 stocks, held in seven overlapping cohorts, and a paper-trading book.

This dataset holds only our own outputs and public inputs. It holds no market data: no Alpaca or Yahoo bars, no prices of any kind, no returns and no labels. Market data from Alpaca may not be redistributed, so anyone who wants to re-run the data, features, models or simulations needs their own free Alpaca paper key.

What you can do with it

With no key at all, the project's command line reads this dataset and re-reports its results:

uv run smz outputs pull        # downloads this dataset into data/publish/, no token needed

The model metrics, the simulation results, the latest scores and the paper book's orders are then read from these files, and the dashboard page is built from them. Precision, AUC and the simulations themselves are not recomputed: they need the labels and the prices, which are not published.

Files

Folder	What
models/current/	the model used live: model file, features.json (feature names), meta.json, the strategy settings
models/<model>/	the last fitted model of each model kind and its settings (settings.json)
scores/<model>.parquet	one row per stock and decision: ticker, decision_utc, set, fold, role, score
metrics/<model>.csv	model metric tables: precision, AUC and information coefficient per fold and per year, top-k tables
sims/report.csv, sims/grid/, sims/costs/	the strategy comparison, the grid of strategy settings tried, the trading-cost tables
sims/<strategy>/<set>/	per strategy of ours: metrics.json, daily equity (daily.csv, equity.parquet), trades as weights and values (trades.parquet)
sims/<benchmark>/<set>/metrics.json	benchmark metrics only (SPY, OEF, equal weight): their equity curves are price paths and are not published
ledger/	the paper book: runs, scores, targets, cohorts, orders (side, value, status), positions and equity, without price columns
macro/fred/macro.parquet	the FRED series the features use, with the time each value became known
membership/	S&P 100 members: today's list and the point-in-time history
site/index.html	the dashboard page, also served by the dashboard Space
manifest.json	the files of the last publish and when they were packed

set is the stock universe: sp100_today (today's members) or sp100_pit (the members at each point in time). Times are UTC.

Sources

• Scores, models, metrics, simulation results, ledger: produced by this project.
• FRED series: Federal Reserve Bank of St. Louis, FRED. Most series are public domain; each series used was checked when published.
• S&P 100 membership: built from Wikipedia's "S&P 100" article and its revision history.
• Not included: Alpaca market data (redistribution not allowed) and Yahoo Finance data (personal use only).

Licence

• Our outputs (models, scores, metrics, simulation results, ledger, page): CC BY 4.0.
• Membership lists: CC BY-SA 4.0, as they derive from Wikipedia.
• FRED series: as stated by FRED for each series.

Not financial advice

These are results of a course project. The paper book trades a simulated account; nothing here is a recommendation to buy or sell anything.