Backtest Data Prep

backtest-data-prep

You hand over a universe definition, a date window, and an output directory. The skill emits a clean OHLCV+volume parquet plus a manifest plus an edge-case log. The downstream consumer is a backtester (your Python loop, QuantConnect, vectorbt, zipline-reloaded, etc.), not a human reader.

The skill's value is correctness. Quants spend 80% of their time on data prep, and the most common backtest bugs come from five sources:

1. Survivorship bias. Backtesting against today's top-500 silently excludes 2022-2024 failures (FRC, SVB, BBBY, SI). The remaining cohort outperforms the true historical population because the losers are gone.
2. Look-ahead bias on fundamentals. Using the latest revised consensus instead of point-in-time. The signal contains forward-looking information that wasn't available at the rebalance.
3. Corporate action mis-adjustment. Pre-split prices not adjusted, or adjusted using the wrong factor.
4. Holiday and half-day handling. Missing days treated as errors or filled wrong.
5. Inconsistent calendar alignment. Calendar dates vs trading dates, US vs global market hours.

The skill addresses all five and documents the treatment in the manifest.

When to invoke

• A quant says "I need clean OHLCV for the top 500 from 2022 to today for my mean-reversion backtest"
• A researcher needs a survivorship-clean dataset for a paper
• A factor-research extension wants a longer window than the in-memory panel can carry
• The user says "prep a backtest dataset", "build me a clean OHLCV parquet", "I need point-in-time prices for a universe"

What you need

• A universe (CLI flag: top100, top500, top1000, sp500, custom:path/to/tickers.csv)
• A window (--window YYYY-MM-DD..YYYY-MM-DD)
• An output directory (--out path/to/dataset/)
• MASSIVE_API_KEY exported in the environment
• Stocks Starter plan minimum (flat files included with any paid plan; flat-files entitlement probed and REST fallback used when not granted)

Expected runtime

A 4-year top-100 run:

• Flat-files cold: ~5-10 minutes
• REST grouped cold: ~10-15 minutes
• Warm (parquet day-cache hit): under 60 seconds

A 5-year top-500 run is roughly 5x the REST time and 2-3x the flat-files time (the bottleneck shifts from day-bucket fetch to per-ticker corp action calls).

What you get back

Three files in the output directory, plus a rendered summary printed to stdout.

Files:

• ohlcv.parquet: one row per ticker x trading day. Columns documented in references/output-formats.md.
• manifest.md: human-readable run record. Universe definition, window, survivorship treatment, corp action methodology, source endpoints, every parameter that affects reproducibility, run timestamp.
• edge-cases.log: line-delimited JSON. One entry per edge case detected: trading halt, IPO partial coverage, delisting during window, ticker change, data gap.

Layer 1: canonical JSON matching output-schema.json. Emitted to stdout alongside the rendered summary so downstream tools that wrap the CLI can parse the run metadata without re-reading the manifest.

Layer 2: rendered dataset summary in the style of a clean-data report (think QuantConnect's data quality report or Bloomberg's BDH audit). See references/rendering.md. Files written, universe construction, corporate actions applied, coverage, edge cases, schema, sources, take. The take answers the question a quant asks at the top of every backtest: "is this dataset actually clean enough to trust?"

How it works

1. Construct the universe per references/survivorship-handling.md. For the default top-N seeds, pull /v3/reference/tickers (active and delisted both), enrich with market cap from /v3/reference/tickers/{ticker}, rank, and keep the top N. The universe label records whether the seed is forward-looking biased (it usually is for current top-N).
2. Filter to common stock. Reference-endpoint type field; drop ETFs, ETNs, ADRCs, units, warrants, rights. Documented in the manifest because the consumer may want to override.
3. Pull daily aggregates for the window. Flat-files preferred (s3://flatfiles/us_stocks_sip/day_aggs_v1/{yyyy}/{mm}/{yyyy-mm-dd}.csv.gz, parallelized 16 workers). REST fallback when entitlement is missing: GET /v2/aggs/grouped/locale/us/market/stocks/{date}?adjusted=true, one call per trading day. Pattern matches factor-research. See ../massive-flat-files/SKILL.md entitlement section.
4. Pull splits and dividends per references/corporate-action-adjustment.md. GET /v3/reference/splits?ticker={T} and GET /v3/reference/dividends?ticker={T}. Massive's grouped aggs are split-adjusted by default (adjusted=true), so the skill does NOT re-apply the split ratio to OHLC; it does emit a cumulative adjustment factor as a separate column so the consumer can un-adjust if their backtester wants raw prints.
5. Pull ticker reference for sector enrichment. One call per ticker, parallelized. Adds sic_code and sector columns to the parquet.
6. Detect edge cases. IPO partial coverage (window starts before list_date), delisting during window, ticker changes (the reference endpoint doesn't link old to new, so flag is best-effort), and trading halts (LULD pauses surface as session OHLC with continuous-trade range; not always detectable from daily aggs). Holidays and half-days are NOT edge cases; they're handled by the calendar (see references/calendar-alignment.md).
7. Emit parquet, manifest, edge-cases log. Parquet uses Snappy compression by default (the right tradeoff for backtest IO; see references/output-formats.md). Manifest is markdown so it diffs cleanly in git.

Foundations used

• massive-flat-files for the bulk daily aggregates pull (S3 auth, path layout, parallelism).
• massive-api-patterns for REST auth, the grouped-daily fallback, and the throttle pattern when pulling per-ticker splits/dividends.

Output mode: dataset

Dataset mode is new in this repo (this is the fifth output mode, joining note, stream, table, exception-report). The defining property: the primary artifact is a machine-readable file on disk, not a rendered string. The rendered string is the operator-readable companion that describes what was written.

The format follows the convention QuantConnect and Bloomberg BDH have converged on for data quality reports: file inventory, universe construction, corporate actions, coverage, edge cases, schema, sources, take. See references/rendering.md for the canonical format rules.

Endpoints used

• s3://flatfiles/us_stocks_sip/day_aggs_v1/{yyyy}/{mm}/{yyyy-mm-dd}.csv.gz: one file per trading day; ~250 files per year. Parallelize 16 workers. Flat-files first.
• GET /v2/aggs/grouped/locale/us/market/stocks/{date}?adjusted=true: REST fallback when flat-files returns 403. Same one-call-per-day pattern.
• GET /v3/reference/tickers?market=stocks&active=true: paginated candidate pool for top-N seeds.
• GET /v3/reference/tickers?market=stocks&active=false: delisted pool for survivorship-clean universes.
• GET /v3/reference/tickers/{ticker}: market_cap (for ranking), type, sic_code, sic_description, list_date, delisted_utc.
• GET /v3/reference/splits?ticker={T}: split history for the corp action adjustment factor.
• GET /v3/reference/dividends?ticker={T}: dividend history for the price-only adjustment factor (when --adjustment total-return).

Doesn't handle (yet)

• Fundamentals join. This skill emits OHLCV only. For fundamentals joined point-in-time, use earnings-drilldown (which carries the 8-K acceptance methodology) or chain factor-research downstream. The schema reserves fundamentals_path for a future fundamentals file in the same output directory.
• Total-return adjustment. Default adjustment is price-only (splits applied by Massive, cumulative factor for un-adjustment). Total-return adjustment (treating dividends as reinvested) is documented in references/corporate-action-adjustment.md but the v1 implementation uses price-only. The adjustment field in the manifest records the choice so the consumer knows.
• Intraday data. Daily aggregates only. Minute and tick aggregates are queued as a separate skill (the file sizes change the problem).
• Spinoff cost-basis allocation. Detected and flagged (spinoffs[] in the schema), but the basis split is not applied; the manifest recommends manual override for spinoff-heavy windows.
• Non-US universes. US common stock only in v1. Massive's flat files cover other locales but the survivorship machinery here is US-specific (CIK linkage, exchange filter, holiday calendar).
• Point-in-time universe reconstruction. Same caveat as factor-research. The default top100/top500/top1000 seeds use current market cap; for a true point-in-time universe, reconstruct per period. Queued as a clean PR extension.

These are clean PR extensions. The output schema reserves space for each so adding them later doesn't break consumers.