Germany MomentumXETRA strategy report

A momentum strategy on Deutsche Börse XETRA, built with the same lookahead-free engine and data-quality discipline as our Canada and Taiwan strategies. Now running as a live beta signal, and subjected to two independent hostile audits (Section IX).

I. Headline Result

II. Data Quality: US ADRs Mistagged as German

EODHD's CountryName field on XETRA tags every US company cross-listed there — Apple, Microsoft, Tesla, Alphabet, Meta, Nvidia, Amazon, Berkshire Hathaway and more — as "Germany", the same mistagging class as Canada's CDR problem. The reliable signal instead: ISIN country prefix (US... vs DE...), which correctly identifies domicile regardless of what exchange metadata claims. A full scan found 226 US-ISIN tickers on XETRA (vs 622 genuine DE-ISIN names, plus another few hundred genuine listings from Switzerland, UK, Austria, Netherlands, France and other European markets XETRA also carries). All 226 excluded per explicit request to test the strategy without US ADRs.

III. A Second Bug: Weekend Contamination in the Signal Itself

The raw union-of-all-tickers price panel carries 1,146 weekend "ghost" dates (average coverage 28 of 1,362 tickers on those dates, vs 465 on real weekdays) — the same calendar-contamination class found in the Canada research earlier this session. There, the bug lived in the execution-lag lookup and was already fixed by sourcing a clean index calendar. Here we found a second instance of the same root cause: the momentum signal itself, computed via the shared backtest engine's own eom = prices.resample(...).last() step, reads directly from this contaminated panel.

IV. Sequential Optimization (Phases 1–3)

One-at-a-time sweeps, same method as Canada/Taiwan: momentum × mcap first (no regime, no filter), then regime MA window fixed on the winner, then fundamental filter fixed on that.

Phase	Winner	Sharpe	MaxDD
1 — Momentum × mcap	9M/skip0, mcap top60%	0.87	−47.7%
2 — Regime MA (fixed on Phase 1)	MA200	0.84	−29.2%
3 — Fundamental filter (fixed on Phase 2)	None	0.84	−29.2%

Unlike Canada (MA75) and Taiwan (MA75), Germany's regime sweep favors the longest window tested — Sharpe rises roughly monotonically from MA50 (0.66) through MA200 (0.84). The fundamental-filter sweep at this stage found EBIT>0 gave by far the best MaxDD (−19.4%) at real cost to Sharpe and CAGR — flagged here because Section VII's joint search later found a materially better version of that same trade-off.

V. Joint Grid — Sequential Wasn't the Optimum

A one-at-a-time search can miss parameter interactions. A reduced joint grid (5 lookback/skip × 3 mcap × 2 MA × 4 N × 2 filter = 240 combinations, run around the sequential winner) found a better config on every axis tested:

Config	Sharpe	Sortino	CAGR	MaxDD
LB12/SK0, mcap top70%, MA200, N=20, None (final)	0.945	1.81	+20.4%	−28.2%
LB9/SK0, mcap top70%, MA200, N=20, EBIT>0	0.899	1.70	+15.0%	−18.0%
Sequential result (Section IV, N=15)	0.843	1.76	+21.9%	−29.2%

VI. Deep Audit of the Final Config

Verified directly against the real backtest (not a proxy): the top 20 most extreme realized ticker-returns the engine actually used top out at +164% (VUL, Jan 2021) — a plausible small-cap momentum spike, nothing resembling the Taiwan/Canada sentinel-value corruption. Two events exceeded the ±90%/300% flag threshold (ONK −92% May 2008, AMC −95% Jul 2008); both fell in regime-defensive months and never reached the published return series (ret = lag_rets × regime_at_signal zeroes the whole month when regime is off). No further exclusions were needed for this final config.

VII. The Alternative Not Chosen

LB9/SK0, mcap top70%, MA200, N=20, EBIT(TTM)>0 — Sharpe 0.899, Sortino 1.70, CAGR +15.0%, MaxDD −18.0%. Also independently verified clean: 0 flagged implausible-return events in this config's real held-position history. This is arguably the stronger risk-adjusted choice (10pp shallower MaxDD for 0.05 less Sharpe) — not used here because CAGR was prioritized for this report, not because it tested worse.

IX. Hostile Audits — What Survived and What Didn't

The strategy was put through two adversarial audits whose explicit brief was to falsify the results, not improve them: a first full audit (independent reimplementation, lookahead hunt, survivorship forensics, placebo battery), and a second audit run without trusting the first. Everything below was measured against the real locked backtest, reproduced first to 10⁻¹⁶ precision.

What survived

• No lookahead, no arithmetic errors. An independent reimplementation with its own signal and eligibility conventions lands at CAGR 19.5% / Sharpe 0.90 vs the engine's 20.4% / 0.95 — all divergences trace to eligibility conventions, none to information leakage.
• Conservative conventions, not favorable ones. Cash earns 0% in defensive months (crediting rf=3% raises Sharpe to 1.01); the DAX benchmark is a total-return performance index, so the +15.4pp alpha is already measured against dividends-included returns; rf=3% is already inside the Sharpe.
• Placebos fail as they should. Inverted ranking (bottom-20 momentum) collapses to Sharpe 0.04; 30 random portfolios through the identical pipeline average Sharpe 0.13 with a max of 0.47 — the real strategy beats the best of all 30. A 1-month-stale signal still earns 17.9%/0.83, the slow-decay signature of genuine 12-month momentum.
• No dependence on outliers or single names. Removing the return-sanity clip entirely changes nothing (0.00pp); removing the top-5 contributing stocks leaves 18.4%/0.84; regime timing only protects drawdown (without it: 20.0% CAGR but −52% MaxDD) — the alpha is the stock selection, not the timing.
• The clean, verifiable window matches the headline. 2015–2026 standalone: +19.6% / 0.96. 2017–2026: +19.3% / 0.92. The published number does not depend on the unverifiable early years.

What didn't

Delisting coverage was then verified externally: 54 documented German delistings 2015–2026 (squeeze-outs, takeovers, pure delistings, insolvencies — Wirecard, Audi, Osram down to small-caps like Aves One, Stemmer Imaging, USU) were cross-checked against the dataset's price-history end dates. 49 of 54 (91%) are present and end in the correct year; the 5 genuinely missing all cluster in 2015–2017 (Sky Deutschland, GAGFAH, Postbank, DVB Bank, McKesson Europe). Conclusion: delisting coverage is effectively complete from 2018 onward, partial 2015–2017, zero before — which is exactly why the 2015+/2017+ windows above are the trustworthy ones.

Three smaller mechanical defects were found and quantified — all with net effect ≈ 0 to +0.3pp when corrected, so none changes the conclusion:

• Duplicate share lines of the same company (BMW/BMW3, Vestas VWS/VWSB, Sixt, Ericsson…) can be held simultaneously — 230 ticker-months of accidental double weight. Deduplicating: 20.7%/0.96.
• The US-ADR exclusion has a blind spot: 464 tickers carry no ISIN, and 8 US names slipped through by that route (American Express, Union Pacific, Parker-Hannifin…), 1.9% of held ticker-months. Removing them: 20.7%/0.96.
• "Death = money back": 34 held positions whose price vanishes mid-month get a 0% return instead of a loss. If those were all −50/−100%: roughly −0.25 to −0.5pp CAGR.

Real measured costs and capacity

Live bid/ask spreads were pulled from Interactive Brokers for the entire 288-name eligible universe (283 quoted): median 35bps, mean 80bps; for the 20 currently-held names, median 22bps with a heavy tail (two names at 118 and 245bps). The 40bps cost assumption is realistic at small size. Stress tests: 100bps round-trip → 18.1%/0.82; 200bps → 14.4%/0.62. Re-running the backtest with today's real spreads as an eligibility filter: names ≤100bps → 19.9%/0.93; ≤60bps → 19.4%/0.90; ≤30bps → 17.9%/0.84. Median position ADV is ~€0.4M/day: treat this as a sub-€1–2M strategy; it does not scale.

X. Limitations

1. Pre-2015 results are unverifiable (survivorship-inflated by construction; see Section IX). Anchor expectations on 2015–2026: +19.6%/0.96.
2. No out-of-sample holdout. Every parameter (Sections IV–V) was chosen looking at the same full 2000–2026 history used to report performance; Germany was also selected from a multi-market search — winner's-curse discount applies on top of the audit's numbers.
3. Capacity is hard-limited. Median position ADV ~€0.4M/day, real spread tail up to 245bps — sub-€1–2M only, limit orders mandatory.
4. Joint grid was reduced, not exhaustive. 240 combinations around the sequential winner, not a full search of the parameter space.
5. Non-US European cross-listings were left in the universe per the design choice; the audit found the related duplicate-line and no-ISIN defects listed above (net effect ≈ 0).
6. Spreads were measured on one day (2026-08-25); depth and market impact were not measured, and historical spreads are unknowable.