loading…
The candidate was tested — walk-forward
loading…
Hypothesis board —
Every hypothesis gets one honest verdict. Lifecycle: idea → proposed →
pre-registered test → verdict. Confirmed = survived the tests; historical = was real,
decayed; candidate = in-sample, awaiting walk-forward; underpowered = right shape, too
little data; rejected = noise/dead. Click a row to jump to the evidence.
Price with moving averages
SMA20 (green), SMA50 (blue), SMA200 (red). Drag the range slider.
Price action around the lunar cycle
Mean daily return per moon phase octile (0 = new moon, 0.5 = full).
Error bars = ±1 SE. The strongest cell overall sits in the trend-conditioned family.
Mean daily return per sun octile (seasonal)
Sun longitude octiles — i.e. time-of-year buckets (0 = spring equinox).
Volatility per moon octile
Mean |return| — moon effects often show up in vol, not direction.
Per-pair overview — every planet combination
Sun–Mars, Sun–Venus, Sun–Mercury, Sun–Moon… then Venus–Mars, Venus–Mercury…
all 45 pairs. Each cell = mean daily return of this instrument on days that aspect is within orb.
Red = flagged slow (<30 episodes — aspect lasts months/years, untestable on daily data).
Per-planet test — add one planet at a time
Each planet gets its OWN family: its aspects to the sun, moon and all other
planets (5 aspects × 9 targets, own 10k block-permutation null on max |cell mean|) — the
incremental path: add Mercury, look at Mercury, then Venus … The right column is the planet's
best aspect-proximity pair (continuous version, 8 octiles from exact-aspect to between-aspects;
pairs with <30 episodes are flagged slow and untestable on daily data).
Aspect proximity — all 45 pairs
Permutation p per pair (block null, 10k). Red = flagged slow (<30 episodes —
few long aspect runs, not testable at daily granularity). Dashed line = 0.05. Expect ~2.25 pairs
below the line by chance alone.
All pre-registered families
Null per family: permutation of the max |cell mean| (10k shuffles) —
the multiple-comparison problem is priced in per family. half-corr = correlation of
cell means between 1975–2013 and 2013–now (cells with ≥50 days in both halves).
Volume & RSI
Payday effect — the salary hypothesis
Hypothesis pre-registered: payday-adjacent days (last + first of month)
show positive excess returns. US pays on the 1st & 15th, Europe mostly at month-end.
On S&P the first-3 days are the strongest calendar effect in this whole system
(p=0.0000) and it existed in every era since 1928. Gold shows a month-end effect
(last-3, p=0.0012) but it is era-dependent, not robust.
Event-adjusted + combinations — the trader's lens
Dirk's points: Gann lived before our sample — known mega-events (1980
bubble, 1987, 2000, 2008, 2020) can contaminate a window; and how would a trader read it
(win rate, net bp after costs — not just p)? Event-adjusted: the new-moon S&P effect
survives (net +6.4bp, p=0.010 without the mega-events); double-ingress is vol-only, net
negative. Combinations: the best combo (new-moon × first-3 × above-MA200, win 61%, net
+17bp) is BEATEN by the control without astrology (first-3 × above-MA200, win 60%, net
+18.7bp) — the astro layer adds nothing on top of calendar+trend.
The metric battery — every hypothesis, five metrics
Family re-test complete (2026-08-16): all bucket families (octiles, 45
aspect pairs, houses, Vedic/BaZi, retrogrades, trend×moon, volume×moon) re-run on all 5
metrics. No new signals — the vol hits decompose into slow-pair artifacts (excluded),
the known below-MA200 vol fact, seasonal aliases, and one real candidate already on the
board (Mars-Jupiter conjunction vol, S&P, 45 episodes, p=0.0044). The astrology picture
is unchanged after the full multi-metric re-test.
The harness (2026-08-08): throw a hypothesis mask in, get returns /
volatility / tops-clustering / bottoms-clustering / asymmetry out. Retroactive sweep of 31
features × 2 markets. Sanity check passed: all known-real calendar effects re-found
(first-3 days p=0.0000, TOM, mid-15, December…). New from the battery: Mars-Jupiter
conjunction vol (S&P p=0.0016), pre-holiday S&P bottoms clustering (p=0.0046 — the
underpowered effect, new metric), new-moon S&P returns (p=0.0052). Re-confirmed:
double-ingress vol, Venus-retrograde bottoms, Mars-Saturn square bottoms.
Gann time cycles — the videos tested
Gann preferences (Dirk's 2nd summary): favored pairs (Sun-Mercury
'bullish', Venus-Jupiter, Mars-Jupiter…) — ~4/60 hits, inconsistent directions, Sun-Mercury
null both markets = noise. Saturn-Uranus square → crashes: anti-finding (6% vs 81% expected,
date-coincidence artifact; the 2000 example was the opposition). Hard/soft + grand
cross/trine: null. The only live lead remains the double-ingress volatility above.
Gann astro pack: planet-in-sign noise (p=0.24/0.34);
price-to-degree scattered (mercury-2° S&P −16.7bp p=0.006, others different planets per
market = noise signature); eclipses + confluence all null; ONE live lead — double-ingress
days (2+ planets enter a new sign) have +7% elevated volatility on both markets, p≈0.000,
survives the calendar-week confound. Volatility, not returns; mechanism unknown — candidate.
The Gann deep-dive (Dirk): the mechanical core tested properly.
G1 calendar days (Gann used calendar, not trading days): 2 of 16 marginal (S&P
+45d p=0.038, +225d p=0.014), no cross-market overlap — noise signature.
G2 anniversaries: p=0.53/0.46 — nothing. G3 zodiac cardinal clustering:
no clustering (S&P actually below random). G4 lunar clustering: turning points
occur at half the random rate near new/full moons — the claim inverted, most likely a
detection artifact. His self-reported track record ("286 trades, 264 wins") is
unverifiable; the Square of 9 is unfalsifiable as published. The mechanical core does
not contain the edge.
Crash autopsy — hypotheses read from history
Dirk's flip: no more proposed hypotheses — the machine reads what the
price did in the 5 days to 5 years BEFORE every crash (S&P: ≥15% drawdown from 1-yr
high, 78 events; gold: ≥20%, 24 events) and tests each feature against random dates.
Found: gold crashes follow gold manias — prior 5y return +95% vs random, p=0.0073.
Falsified: extended S&P markets before crashes (p=0.97 — crashes come from all states),
and gold as a reliable equity safe haven (flat +0.1% during crashes).
Practitioner pack — what astro traders ACTUALLY use
The real-world claims, tested with the RIGHT metric. Mercury retrograde is
used as a volatility filter — so we tested volatility: no effect (S&P retro vol 77.0
vs 76.2, gold 78.3 vs 82.2; shadow windows and stations all null). Outer-planet stations:
no vol spikes. Mars/Venus hard aspects to Saturn/Uranus: no effect on returns or vol. The
practice layer is as dead as the theory layer — on 3,110 station days and 4,758 retro days.
Market structure — the overnight effect
Split of the daily return into Overnight (prev close → open) and Intraday
(open → close). Gold: the return is an overnight phenomenon (p=0.0012) — 78–103% of it
happens between close and open; the session itself is flat. S&P: the famous overnight
effect has decayed — the intraday session contributes again since 2013. The overnight
premium is risk compensation (the gaps live there), not a free edge.
Telegram-kanaal audit — EmperorBTC vs Wimar.X (19-08-2026)
Beide kanalen causaal gevolgd met onze eigen regels (entry op post-moment,
SL 2×ATR, TP R:R 3:1, kosten): EmperorBTC (analysis-kanaal): −2,5% over 9 signalen.
Wimar.X (nieuws/trendvolgend): 58 headline-signalen, €1000 → €1.472 (+47%, maxDD 9%),
p=0,005 vs random — terwijl BTC zelf −33% deed. De les: trendvolging wint, gepraat
niet. Wimar.X is geen voorspeller — het volgt de trend, en dat doen wij met de EMA-basis
al zonder de clickbait. Forward-test (woordenlijst pre-registered) loopt.
TA-basis — het fundament (Dirk 2026-08-19)
De trading-basis die ALTIJD in place is: trend (EMA50/200, prijs>MA200),
stop loss = invalidation (ATR×2), R:R 3:1, 1% risico, kosten+slippage. €100-start simulaties
per risicoprofiel (laag 0,5% / middel 1% / hoog 2,5%), volledig causaal. Goud pre-TA
(1975-90) is het beste tijdperk — 58% win vs 42-44% later (minder arbitrage). S&P 30jr:
model 1,24× B&H met 4,7× minder drawdown. Crashes: model wint of beschermt 80-164%. De
manual-strategieën (RSI-divergentie, volume, candlestick) zijn eerste runs — null-test volgt.
Alles wat we bijbouwen (astrologie, exotische data) bouwt BOVENOP deze basis.
Predictable events — scheduled + geopolitical
FOMC decision days (dates published years ahead) + geopolitical shocks.
Confirmed: S&P volatility is higher on FOMC days (p=0.022); gold +2.4% and S&P
−1.7% in the 5 days after a major geopolitical shock (p=0.009 / p=0.019). Shocks are
not predictable in advance — but the reaction pattern is real and actionable after the fact.
Selling dates — when people liquidate savings/gold
Pre-registered: quarterly US tax dates (15 jan/apr/jun/sep), the 15th,
December tax-loss window, month-end, quarter-end. Finding: gold was weak on the 15th
for 38 years (−16.8bp/day, 1975–2013) and the effect completely stopped after 2013
(historical). December-end is POSITIVE, not negative (Santa rally > tax-loss selling).
The month-end drain on S&P appears only in the modern era (−3.8bp 2000+).
Spending seasonality — when people MUST pay
Pre-registered windows: January (NL insurance year-start, US deductibles
reset), May (NL vakantiegeld), gold autumn-winter (CNY/Diwali demand), Halloween (S&P).
Finding: gold January survives out-of-sample (half2 p=0.0067, +19.9bp/day over 13 years)
— first calendar candidate since payday. S&P January decayed after publication
(+5.2bp pre-1990 → −1.8 now). May/vakantiegeld: nothing.
Earnings season — the quarterly results hypothesis
Pre-registered calendar proxy: the 20 trading days from the 3rd Monday of
Jan/Apr/Jul/Oct (when US quarterly results dominate) + quad witching (3rd Friday).
Result: no return or volatility effect on S&P or gold with this proxy (S&P
earnings season p=0.86). Real announcement-day effects need an actual earnings calendar
(API key), not a calendar window.
For comparison — calendar effects that ARE real
The same test machinery on S&P 500. Turn-of-month is the textbook real
effect: p=0.0000 — zero of 10,000 shuffles matched it, and both halves agree. Pre-holiday
shows the right shape but is underpowered here (only 282 days). The astro cell looks
"significant" in-sample — and fails half-2 (p=0.17). In-sample p is cheap; the walk-forward
is the judge.
Event windows (directional pre-registered tests)
Method — written before results were seen
| Rules | |
| Data | |
| Ephemeris | Skyfield DE440, tropical geocentric-apparent, 00:00 UTC per trading day; outer planets via barycenters (<0.01° error, below all orbs) |