Degrees of Freedom in Lunar Financial Backtests
How Window Tuning, Overlapping Horizons, and Bid–Ask Spreads Manufacture Syzygy Return Anomalies
- Literature Baseline
- Dichev & Janes (2003); Yuan, Zheng, & Zhu (2006)
- Multiple-Testing Hurdle
- |t| > 3.0 (Harvey, Liu, & Zhu, 2016) + BH-FDR q-values
- Autocorrelation Fix
- Event-clustered / Newey–West HAC standard errors
- Data-Rights Posture
- Raw third-party market series WITHHELD by release lockfile
Give an unconstrained researcher seven window widths, four asset classes, equal-weighted microcaps, and zero transaction costs, and a periodic 29.53-day cycle will clear p < 0.05 by degrees of freedom alone.
Evidentiary Scope: Methodological audit of published 'New Moon minus Full Moon' equity return papers against multiple-testing t-statistic hurdles (Harvey, Liu, & Zhu, 2016), overlapping-window autocorrelation, and realistic execution costs.
1. The garden of forking paths around a 29.53-day cycle
At first glance, testing whether stock returns differ between New Moon and Full Moon sounds like a single, clean binary hypothesis. In actual econometric code, a researcher immediately confronts at least five unstandardized specification choices: do you score the exact UTC syzygy date, ±1 trading day, ±3 calendar days, ±7 calendar days, or a continuous cosine wave cos(Δλ)? Do you align UTC oppositions to New York 16:00 close, Tokyo close, or next-day open? Do you weight index constituents equally or by market capitalization?
In empirical asset pricing, Harvey, Liu, and Zhu (2016) demonstrated across hundreds of published factor claims that conventional |t| > 1.96 thresholds severely understate false-discovery rates once researcher specification search is accounted for—recommending |t| > 3.0 alongside explicit family-wide false-discovery controls. Robert Novy-Marx (2014) drove the point home with biting clarity by showing how persistent or cyclical non-economic series (global weather, sunspots, and planetary configurations) can generate seductive backtest t-statistics whenever specification degrees of freedom go unpenalized.
| Specification Axis | Permissive Backtest Choice | Statistical / Economic Distortion | Observatory Methodological Rule |
|---|---|---|---|
| Phase Window Width | Trying ±1d, ±3d, ±7d, ±15d and reporting the lowest p-value | Multiplies effective hypothesis count without family adjustment | Pre-declare window & apply Benjamini–Hochberg FDR across grid |
| Multi-Day Horizon Overlap | Treating 7 consecutive window days as 7 independent draws | Autocorrelated volatility shocks inflate naive OLS t-stats | Cluster errors by distinct lunation / Newey–West HAC lags |
| Portfolio Weighting | Equal-weighted international or microcap indices | Concentrates signal in illiquid small caps with wide spreads | Separate large-cap value-weighted indices explicitly |
| Execution Friction | Zero-cost close-to-close paper switching 24× per year | 24 round-trip turnovers (~10–25 bps each) erase ~3–5%/yr spread | Evaluate net of realistic bid-ask + slippage bounds |
| Sub-Period Stability | Full-sample 1970–2000 average dominated by pre-decimalization eras | Effect attenuates sharply in post-2000 institutional markets | Report decade-by-decade out-of-sample decay |
Synthesized from Yuan, Zheng, & Zhu (2006), Novy-Marx (2014), and Harvey, Liu, & Zhu (2016). Not investment advice.
2. Signal-to-noise ratio and round-trip turnover drag
Even taking Yuan, Zheng, and Zhu’s (2006) 48-country equal-weighted estimates at face value—roughly 3% to 5% annualized return difference between 7-day/15-day New Moon and Full Moon windows across 1973–2001—divide that figure by 12.37 synodic lunations per year and the per-cycle differential sits around 25 to 40 basis points, buried inside daily equity standard deviations of 100+ basis points.
Notice also that Yuan et al. themselves reported substantially weaker and statistically insignificant spreads when shifting from equal-weighted portfolios to value-weighted large-cap portfolios, where institutional arbitrageurs dominate and bid–ask spreads are tight. Switching an equity book between risk-on and cash ~24 times a year incurs spread, market-impact, and tax drag that routinely exceeds 25 basis points per cycle in the very small-cap stocks driving the equal-weighted average.
3. Why Lunologium enforces a cryptographic release lock on market files
Because celestial market screens invite confirmation bias—and because commercial historical market feeds prohibit raw public redistribution—Lunologium separates public methodology from restricted market artifacts via an explicit SHA-256 release gate (scripts/verify-data-rights.mjs).
Visitors can audit our exact clustering formulas, Benjamini–Hochberg false-discovery procedures, and peer-reviewed literature summaries openly at /markets/methodology and /markets/does-the-moon-affect-stock-market, while our build pipeline refuses to deploy if any unapproved market snapshot enters public/api or src/data.
Companion Observatory Calculators & Ledgers
Literature review and release-hold documentation.
Clustering, false-discovery controls, and uncertainty bounds.
Parallel cohort-power and replication audit in human sleep biology.