alansynn/scots

Data Manifest

Updated: 2026-03-19 | Parquet-first data flow (bridge store-free manager)

Parquet Datasets

Dataset	Layout	Symbols	Range	Use
parquet/5min/crypto_bybit/ohlcv/	year/month partitioned parquet	major crypto (BTC/ETH/SOL...)	requested from 2018-01-01	Bybit 5min OHLCV
parquet/5min/crypto_bybit_futures/ohlcv/	year/month partitioned parquet	major crypto futures (BTC/ETH/SOL...)	requested from 2018-01-01	Bybit futures 5min OHLCV
parquet/1min/crypto_coinbase/ohlcv/	year/month partitioned parquet	BTC/ETH/SOL/XRP/ADA/LTC/DOGE/BCH	requested from 2018-01-01 (listing-aware)	Coinbase 1min OHLCV
parquet/1min/crypto_binance_us/ohlcv/	year/month partitioned parquet	BTC/ETH/SOL/XRP/ADA/LTC/DOGE/BCH	requested from 2018-01-01 (first-bar auto-detect)	Binance.US 1min OHLCV
parquet/1min/crypto_us_merged/ohlcv/	year/month partitioned parquet	merged US crypto venue set	rolling	Coinbase + Binance.US merged 1min
parquet/5min/us/ohlcv/	year/month partitioned parquet overlay	US universe	rebuilt from local IEX + us_merged fallback	repaired 5min US runtime path
parquet/5min/us_yahoo_merged/ohlcv/	year/month partitioned parquet	US universe	rolling	US 5min backtest
parquet/daily/us/ohlcv.parquet	single parquet	US universe	long history	US daily backtest
parquet/universe/*.parquet	single parquet per universe	index members	historical	universe filtering

Data Management

All data operations go through the unified scripts/data_manager.py:

uv run python scripts/data_manager.py status     # parquet freshness, row/symbol/range
uv run python scripts/data_manager.py health     # Stale symbols, quality checks
uv run python scripts/data_manager.py update --db bybit_crypto_5min --start-date 2018-01-01
uv run python scripts/data_manager.py update --db bybit_crypto_futures_5min --start-date 2018-01-01
uv run python scripts/data_manager.py update --db coinbase_crypto_1min --start-date 2018-01-01
uv run python scripts/data_manager.py update --db binance_us_crypto_1min --start-date 2018-01-01
uv run python scripts/data_manager.py merge --profile crypto_us_1min_merged --start-date 2018-01-01 --end-date 2026-12-31
# Large windows: split by year and use --append after first rebuild
uv run python scripts/data_manager.py merge --profile crypto_us_1min_merged --start-date 2018-01-01 --end-date 2018-12-31
uv run python scripts/data_manager.py merge --profile crypto_us_1min_merged --start-date 2019-01-01 --end-date 2019-12-31 --append
uv run python scripts/data_manager.py reconcile   # parquet integrity checks
uv run python scripts/data_manager.py merge --profile us_repaired --start-date 2018-01-01 --end-date 2026-03-06
uv run python scripts/data_manager.py merge --profile us_yahoo_merged --start-date 2018-01-01 --end-date 2026-03-06

As of 2026-03-19:

• yahoo_daily is refreshed through 2026-03-18.
• yahoo_5min is refreshed through 2026-03-19.
• the us_merged, us, and us_yahoo_merged overlays were append-rebuilt for 2025-10-01 ~ 2026-03-19.
• the canonical promotion-grade US 5min parquet aliases now reach 2026-03-19 and recover to 533/719 active symbols fresh within 7 days (74.1%), but their dataset-wide positive_volume_ratio remains only about 0.0928 ~ 0.0930, still below the 0.20 fail-closed threshold.

Full-history US 5min promotion therefore remains intentionally unsupported on the current local source stack. Use config/backtest/champions/5min/REGISTRY.md and docs/PROJECT_STATUS.md as the source of truth.

data_manager.py는 bridge store 경로를 사용하지 않고 parquet-only로 동작합니다.

Bulk crypto parquet backfill helper:

data/scripts/crypto_parquet_backfill_all.sh

Parquet-First Sources

Sources are parquet-native and do not use bridge store as an intermediate store.

• bybit_crypto_5min:

  • updater: data/scripts/collect_5min_bybit_parquet.py
  • output: data/parquet/5min/crypto_bybit/ohlcv/year=YYYY/month=MM/<SYMBOL>.parquet
  • data_manager.py update --db bybit_crypto_5min: collect + write parquet directly
  • data_manager.py export --db bybit_crypto_5min: no-op (already parquet-first)
  • resume 안정성: manifest 기반 완결성 검증(_ingest_manifest.jsonl) + 필요 시 재수집
  • 소스 전략:
    • primary: kline_for_metatrader4 5분 OHLCV (일부 심볼/월은 2024-11까지)
    • fallback: public.bybit.com/spot 일별 체결 파일 → 5분 OHLCV 집계 (누락/부분 월 보완)
  • 참고: 일부 실행 환경에서는 Bybit v5 API가 지역 정책으로 403 차단될 수 있어, 공개 파일 기반 fallback 경로를 기본 사용
  • 보조 휴리스틱: --resume-min-ratio (기본 0.98, manifest 누락 시에만 사용)

• bybit_crypto_futures_5min:

  • updater: data/scripts/collect_5min_bybit_futures_parquet.py
  • output: data/parquet/5min/crypto_bybit_futures/ohlcv/year=YYYY/month=MM/<SYMBOL>.parquet
  • source: public.bybit.com/trading/<SYMBOL>/<SYMBOL><YYYY-MM-DD>.csv.gz 일별 체결
  • 수집 방식: 일별 체결 -> 5분봉 OHLCV 집계 -> 월 parquet 저장
  • data_manager.py update --db bybit_crypto_futures_5min: collect + write parquet directly
  • 선물 롱/숏/레버리지 백테스트는 region = "crypto_bybit_futures" 사용을 권장

• coinbase_crypto_1min:

  • updater: data/scripts/collect_1min_coinbase_parquet.py
  • output: data/parquet/1min/crypto_coinbase/ohlcv/year=YYYY/month=MM/<PRODUCT>.parquet
  • source: api.exchange.coinbase.com/products/<PRODUCT>/candles (granularity=60)
  • 수집 방식: 300-candle chunk API 호출 -> 월 parquet 저장
  • 기본 대상: BTC-USD, ETH-USD, SOL-USD, XRP-USD, ADA-USD, LTC-USD, DOGE-USD, BCH-USD
  • 히스토리: 2018-01-01 요청 기준, 심볼 상장일 이전 구간은 자동 skip

• binance_us_crypto_1min:

  • updater: data/scripts/collect_1min_binance_us_parquet.py
  • output: data/parquet/1min/crypto_binance_us/ohlcv/year=YYYY/month=MM/<SYMBOL>.parquet
  • source: api.binance.us/api/v3/klines (interval=1m)
  • 수집 방식: 1000-candle chunk API 호출 -> 월 parquet 저장
  • 기본 대상: BTCUSDT, ETHUSDT, SOLUSDT, XRPUSDT, ADAUSDT, LTCUSDT, DOGEUSDT, BCHUSDT
  • 히스토리: 심볼별 첫 캔들을 자동 조회해 pre-listing 월을 skip

• crypto_us_1min_merged:

  • profile: crypto_us_1min_merged
  • source priority: coinbase_crypto_1min -> binance_us_crypto_1min
  • output: data/parquet/1min/crypto_us_merged/ohlcv/year=YYYY/month=MM/part-0000.parquet

Schema

Daily:

daily_bars(symbol, date, open, high, low, close, volume, adj_close, dividends, splits)

5min:

bars_5min(symbol, timestamp, open, high, low, close, volume, trade_count, vwap)

API Strategy

US Daily Hybrid:

1. DefeatBeta API first (fast, common stocks only)
2. yfinance fallback (ETFs, indices, delisted)

US Intraday: yfinance only (5min data)

KR: KIS API (Real trading credentials required in .env)

KR Market (KOSPI 200)

• Universe: Naver API (https://m.stock.naver.com/api/json/sise/enrollItemListJson.nhn?pageSize=202)
• Data: KIS API, 실전투자 only
• Endpoint: /uapi/domestic-stock/v1/quotations/inquire-time-itemchartprice (tr_id: FHKST03010200)
• Credentials: .env requires KIS_REAL_APP_KEY, KIS_REAL_APP_SECRET

Split Adjustment Strategy

Problem

5min IEX/Tiingo data was collected incrementally — older data was retroactively adjusted at download time, but subsequent splits were NOT applied to already-collected data. This creates price discontinuities at split dates.

Daily Yahoo data is pre-adjusted (adj_close), so daily has no split issues.

Current State (as of 2026-01-31)

• Daily (Yahoo): Fully split-adjusted. No action needed.
• 5min (IEX/Tiingo → Parquet): Fixed via scripts/fix_5min_splits.py. 50 splits detected and corrected (21 forward, 29 reverse). Zero remaining after fix.
• 5min (Yahoo): Short history (60d rolling). Yahoo provides adjusted data. No action needed.
• KR (KIS): No splits detected in KOSPI 200 data.

Scripts

Script	Purpose	When to Run
scripts/fix_5min_splits.py	Detect + fix splits in 5min Parquet	After new data ingestion
data/scripts/detect_splits.py	Detect splits from Yahoo daily (close vs adj_close)	Ad-hoc diagnostics
data/scripts/apply_split_adjustments.py	Apply split adjustments to Tiingo 5min bridge store	Legacy (pre-Parquet)
data/scripts/normalize_tiingo_splits.py	Normalize Tiingo vs Yahoo pricing	Legacy (pre-Parquet)

Ongoing Process

After each 5min data refresh:

# 1. Dry-run: detect splits without modifying
uv run --no-project --with pyarrow --with pandas --with numpy \
  python3 scripts/fix_5min_splits.py --dry-run

# 2. Apply fixes (creates .bak backup automatically)
uv run --no-project --with pyarrow --with pandas --with numpy \
  python3 scripts/fix_5min_splits.py

# 3. Verify: re-run dry-run to confirm 0 splits remain
uv run --no-project --with pyarrow --with pandas --with numpy \
  python3 scripts/fix_5min_splits.py --dry-run

Known Exclusions

Leveraged/inverse ETFs (TQQQ, SQQQ, SOXL, SOXS, UVXY, SPXU, etc.) have frequent reverse splits that are correctly detected and adjusted. Their cross-timeframe close prices may differ significantly from daily Yahoo data due to different adjustment baselines — this is expected and not a data quality issue.

NaN-Gap Handling (Rust Engine)

The 5min sparse matrix has NaN gaps where a symbol has no bar at a given timestamp (e.g., pre-market/after-hours gaps, trading halts). The backtest engine (scots-engine) uses a last-known-good price fallback:

• For held positions with missing price data, the last valid price is substituted
• A summary warning is logged: total_nan_gaps = N
• This prevents phantom equity crashes from the sparse matrix

Data Quality

• DefeatBeta: Fast, common stocks, no rate limits
• yfinance: Slower, supports ALL types (ETFs, indices, delisted)
• KIS: Official broker, rate limited, requires auth

Rate Limits

Source	Limit	Note
DefeatBeta	None	Very fast
yfinance	~2000/hr	100ms delay
KIS Real	20/s	200 symbols = 10s
KIS Paper	2/s	100 symbols = 50s