Skip to content
citrusquant

Playground — data & internals

The interactive playground runs a real backtest in your browser — the yuzu engine and the lemon parser are compiled to WebAssembly, and every run is evaluated against real daily bars for 10 US large-caps (Nov 2014 – Nov 2017). Nothing is sent to a server.

This page is the fine print behind that demo: exactly what the data is, why it ends where it does, and the pipeline each run goes through.

About the data

Section titled “About the data”

The sample dataset is real market data, chosen so it can live in a public repo without a license fight:

  • Prices — daily OHLCV for AAPL, MSFT, NVDA, AMZN, WMT, JPM, GS, XOM, JNJ and PFE, from the Huge Stock Market Dataset (CC0 / public domain), adjusted for splits and dividends.
  • P/E — computed as adjusted close ÷ the last reported fiscal-year diluted EPS from SEC EDGAR XBRL filings (US government data, public domain). Each 10-K’s EPS only becomes visible on its filing date — not the fiscal-period end — so there is no look-ahead. Try is_smallest(pe, 3) and note that AMZN’s P/E goes missing during 2015: its FY2014 EPS was negative, so trailing P/E is undefined.
  • Benchmark — every run is compared against a benchmark series (the engine’s benchmark_key config). When the bundled data ships an index series it is used directly; otherwise the playground builds a daily-rebalanced equal-weight basket of the same 10 names. For a stock-picking strategy on a fixed universe that’s the honest yardstick anyway: it answers “did picking these names beat just holding all of them?” — alpha, beta, excess return and the relative tabs all measure against it.

Why does the data end in 2017? Nearly every “free” market-data source (Yahoo, Stooq, FMP, Tiingo, …) forbids redistribution, which rules them out for a static site. This CC0 dataset is the newest daily OHLCV that is genuinely public domain. The engine itself doesn’t care — it ships no data and runs on any panels you feed it: Bring your own data. Everything is reproducible via fetch-sample-data.mjs.

One honest caveat: bundled prices are dividend-adjusted while EPS is as-reported, so P/E early in the window is understated by a few percent for high-dividend names. Good enough to demo mechanics; bring your own data for research.

How it works

Section titled “How it works”
  1. Your source is parsed by lemon-wasm into a JSON Expr tree (the spec).
  2. The spec, the sample panels, a benchmark series and a config (fee_ratio, benchmark_key, bootstrap_samples) are handed to yuzu-wasm.run_backtest(...).
  3. The returned Report — equity, drawdown and rolling series, calendar returns, bootstrap confidence bands, the full trade list and ~40 metrics — is rendered into the tabs beside the editor. The frontend only draws; every number is computed by the engine (plus a few purely presentational reshapes).

That’s the same pipeline the native engine runs — see Reading a report to decode the output.

Things to try

Section titled “Things to try”

Open the playground and paste any of these into the editor:

  • is_largest(sma(close, 2), 1) — concentrate into a single name.
  • close > sma(close, 50) — a pure trend filter across the whole universe.
  • is_smallest(pe, 3) and (close > sma(close, 20)) — cheap and trending.
  • is_largest(rsi(close, 14), 3) — momentum by RSI. With NVDA’s 2016–17 run in the universe, momentum strategies have an unfair advantage — watch what they hold.

New to the syntax? Start with Your first strategy, or browse the lemon reference for the complete operator set.

Section titled “Shareable links”

Every run has a link. Hit 🔗 Share next to Run and the playground copies a URL with your strategy encoded in it (/playground#s=…) — open it and you land on the same strategy, already loaded. The strategy lives in the URL hash, so it’s never sent to a server; the link is self-contained.

Docs, issues, and posts can deep-link the same way — for example, open these in the playground: