Skip to content

README.md

Latest commit

 

History

History
104 lines (86 loc) · 5.39 KB

README.md

File metadata and controls

104 lines (86 loc) · 5.39 KB
Raw

Research data

Everything the research produced, as plain CSV and Markdown, for anyone who wants to check the results or improve the model. Refresh it from the latest GitHub Actions runs with:

bash scripts/export_research_data.sh

Ground rules used throughout

  • Two periods. 2006-2014 ("1st", "2006_14") is used to choose things; 2015-2026 ("2nd", "2015_26") is only used to check them. Anything chosen by looking at 2015-2026 is labelled as in-sample where it appears.
  • Walk-forward. Every model is refit yearly on the previous 5 years and forecasts the following year only.
  • Costs come from measured Alpaca bid-ask spreads by liquidity tier (studies/spreads*.csv) plus 1 bp of slippage; "break-even" is the one-way cost at which a strategy stops making money, in basis points.
  • IC is the daily cross-sectional rank correlation between a forecast and the next N days' returns (N = horizon_days). An IC of 0.02 is useful.
  • Universe: today's ~1,000 most traded US stocks. This flatters some signals (survivorship bias) — see studies/survivorship.md and the survivorship_prone column.

Where to start

File What it is
zoo/best_correlations.csv The main table. Every signal, signal pair and fitted model, ranked by its IC in the weaker of the two periods.
zoo/README.md Summary of ~550 fitted models: which families work, pros and cons, whether period-1 winners stayed winners, the luck hurdle.
studies/ALL_STUDIES.md All study write-ups in one document.
studies/strategies.md 40+ strategies beyond stock reversal (trend, risk parity, pairs, factors, volatility premium), with correlations.
studies/multi.md The multi-strategy book (ETF sleeves + stock book) as it will trade.
backtest/README.md Full backtest of the stock book (2004-2026) with charts.

zoo/best_correlations.csv columns

Column Meaning
what signal, pair (two signals summed or multiplied) or model
name, family what it is; ideas are tagged idea:<theme> (wave N)
horizon_days forecast horizon: 1, 5, 21 or 63 trading days
ic_2006_14, ic_2015_26, t_* mean daily IC and its t-statistic in each period
ic_2015_20, ic_2021_26 the same split inside 2015-2026, for signals that only exist from 2015 (news)
years_positive share of calendar years with a positive IC
breakeven_bp_*, net_sharpe_* trading it as a long/short book (models and ideas)
same_sign_both_periods, weaker_period_ic the ranking key: did it keep working, and how strongly at worst
survivorship_prone size/illiquidity-type signals that look better on today's survivors than they were

A negative IC that is stable in both periods is just as useful: trade it the other way.

Folders

  • zoo/ — model zoo. zoo_all_strategies.csv has every model x signal set x horizon x smoothing x universe combination (~3,000+ rows) with ICs, Sharpe and break-even costs; parts/ holds the raw per-group results (zoo_<group>_h<horizon>.csv).
  • studies/ — one file set per study:
    • edge* signal ICs, decay and break-even costs; alt* fundamental and news signals
    • combos*, stability.csv signal pairs and which ones kept their sign
    • conditional* when each signal works (regimes, volatility, news days)
    • poisson* jump clustering (INGARCH, Hawkes, Poisson HMM) and jump-based signals
    • hmm* market-regime model variants and calibration
    • strategies*, strategy_returns.csv daily net returns of every strategy; strategy_correlations.csv
    • multi* ETF sleeves + stock book; intraday* intraday strategies
    • events.md, news_events.csv returns after each type of news; survivorship*
    • spreads* measured bid-ask spreads by liquidity tier, time of day and year
  • backtest/ — the reference stock-book backtest (daily.csv: daily returns, exposure and turnover; signal_ic.csv; model_comparison.csv; charts) and all_runs.csv, a one-line summary of every backtest variant tried.

Raw data

The raw inputs are not here, because their licences do not allow redistribution:

  • Daily prices: Yahoo Finance — anyone can download them with python -m qtrade download.
  • 5-minute and 1-minute bars, news: Alpaca (free account) — anyone can rebuild them with their own Alpaca keys: python -m qtrade intraday --timeframe 5Min --symbols 1000 and python -m qtrade news. The repository's own copies are stored encrypted.
  • Macro series: FRED (free API key) or the Yahoo stand-ins in qtrade/data/public.py.

Datasets that can be shared openly — SEC fundamentals, SEC insider transactions and S&P 500 index changes — are published unencrypted to the GitHub release public-data by the publish-public-data workflow.

Reproducing everything

  1. pip install -r requirements.txt
  2. python -m qtrade universe && python -m qtrade download
  3. python -m qtrade.research --tasks edge,combos,strategies,multi (any of quality, spreads, edge, alt, poisson, conditional, combos, garch, glm, strategies, multi, events, survivorship, intraday, hmm)
  4. python -m qtrade.research.zoo --group tree --horizon 5 (groups: simple, linear, tree, nn, extra, ideas, seq, online, targets, features)

Or push a commit with [research: edge,combos] / [zoo: tree@5,21] in its message and GitHub Actions runs it on real data; results land on the results branch.