# Sixty strategies against buy-and-hold: the findings *Written 2026-08-11 at the end of the core research programme in this directory, and updated 2026-08-18 for the entries registered since (the dated addenda at the bottom carry their full detail). The headline counts on this page are generated from the ledger when the site is built, never typed. Every number in this document is in `registry.json` or in a run recorded there — if a figure is not in the ledger, it is not in this write-up.* *Audience: me in six months, and anyone who might otherwise spend a month repeating this. Terms are explained the first time they appear, because I was a beginner when this started and the write-up should not require more than I had.* --- ## 1. The answer **Sixty trading ideas were written down in full before being tested, tested once each against a fixed bar, and one survived.** The count: **1 pass, 16 mixed, 41 fail, 1 invalid, 1 rejected.** The first fifty concluded the core programme on 2026-08-11. Five later entries (2026-08-18, all MIXED) did not change the answer, but two of them sharpened it: the take-profit overlay is now the ledger's cleanest demonstration of the coin-flip signature (§4), and the multi-asset trend rule is the first candidate to fail for a portfolio reason rather than a returns reason (§8). A few terms, once: - A **backtest** replays a trading rule against historical prices to see what it would have earned. Backtests lie easily, which is what most of this document is about. - **Buy-and-hold** means buying the asset once and doing nothing. It is the laziest possible alternative, and it is the opponent every idea here had to beat — because if a rule cannot beat doing nothing, the rule is the wrong place to put effort. - **CAGR** (compound annual growth rate) is growth per year. All CAGRs here are **net of trading costs**, because costs flipped real conclusions in this project — one Vietnam strategy earned +270% before costs and lost everything after them. - A **drawdown** is a fall from a high point to a low point. **Max drawdown** is the worst one — the deepest hole you would have had to sit through without panicking. - **MAR** is CAGR divided by max drawdown: return earned per unit of worst-case pain. - **Sharpe** is return divided by day-to-day wobble (volatility) — a second opinion on the same question, reported so it can disagree with MAR. **The bar.** Every idea was scored on data it had never seen (see §4) and had to beat **every** benchmark — the traded asset's own buy-and-hold, the market index, and an equal-weight basket both rebalanced and left alone — on all three of: MAR at least **1.30×** the benchmark's, max drawdown no worse than **0.75×**, CAGR at least **0.80×**. The margins are the point: "slightly better risk-adjusted return" is what noise looks like. A 1980s calendar effect got MAR 0.50 against the market's 0.46 and was correctly killed by exactly that margin. ### The survivor: `btc-trend-100d` Hold Bitcoin while its price closes more than 2% above its own 100-day average; sell when it closes more than 2% below; do nothing in between. Decided at each daily close, traded at the next open, 10 basis points cost per side, on Binance data (the venue the order would actually rest on). On the test window (2023-11 to 2026-08, never touched during design): | | CAGR | max drawdown | MAR | |---|---|---|---| | the rule | **+36.3%/yr** | **−27.4%** | **1.32** | | just holding Bitcoin | +22.2%/yr | −53.0% | 0.42 | It also worked on five coins it was never designed on (it lost −10.2%/yr there while holding those coins lost −26.8%/yr in a falling market — losing much less than doing nothing is the edge working). **The honest asterisk, and it matters:** `donchian55-btc` — a different classic trend-following construction (buy 55-day highs, sell 20-day lows) on the *same coin, same window, same costs* — **failed** (+14.3%/yr, MAR 0.39). If "trend-following works on Bitcoin" were true as a general claim, both should have passed. One construction passing while its sibling fails means the edge is **specific to this rule on this asset**, not a law of nature, and the two must never be cited as confirming each other. The other early result, `trend3x-150d` (hold a 3× leveraged Nasdaq fund while the Nasdaq is above its 150-day average), is recorded as **MIXED**: it crushes holding the leveraged fund itself (+31.7%/yr at −40.4% vs +17.3%/yr at −81.7%) but its drawdown is *deeper* than simply holding the unleveraged index (−35.1%) or the S&P (−24.5%), and on four sectors it was never designed on it lost −3.8%/yr while holding them made +6.9%/yr. It reduces one specific fund's drawdown; it is not a general edge. --- ## 2. Three attempts to overturn the answer The first 34 ideas (batches 1–2) produced zero survivors. Rather than stopping there, the programme made three deliberate attempts to prove that result wrong. ### Attempt 1: more rules (34 ideas, zero survivors) Every major family of published price-based ideas: calendar effects, mean reversion (buying dips), momentum, volatility filters, trend rules, cross-sectional ranking. Each traced to a published academic source, each written down in full before testing. The consistent killer was the **CAGR gate, not the drawdown gate**: every rule that reduced drawdown did it by being out of the market, and being out of the market costs more return than it saves. That sentence is the single most repeated finding of the whole programme. ### Attempt 2: better data (10 ideas, zero survivors) The sweep's own conclusion was that the binding constraint was the *data*, not the rules: the price cache stored only open/high/low/close, so the strongest published anomalies — volume effects, volatility-index signals, interest-rate regimes, fundamentals — could not even be registered. So the data layer was extended: trading volume for 43 instruments, the VIX family (an index of how expensive stock-market insurance is — high VIX means fear), Treasury and credit interest rates, and company fundamentals from SEC filings. Every new series was integrity-tested before use (§4 and §6 cover what those tests caught). Ten ideas were registered against the new data — the high-volume return premium, volume-conditioned reversal, illiquidity ranking, VIX term-structure timing, the variance risk premium, VIX-spike buying, yield-curve and credit-spread regime gates — each citing its published source. **All ten failed or scored mixed.** The closest single-rule miss of the first fifty was here: `credit-spread-gate-spy` (hold the S&P while corporate borrowing spreads are below their recent median) posted MAR 0.47 vs the market's 0.46 with a little over half the drawdown (−19.0% vs −33.7%) — and still failed, because +8.9%/yr is 0.58× the market's return and the bar is 0.80×. Cutting the hole in half while giving up 42% of the growth is not beating buy-and-hold. ### Attempt 3: combination (3 portfolios, zero survivors) Several failed rules cut drawdown while giving up return — raw material, in theory, for a **portfolio**: combine many mediocre-but-different rules and diversification does the rest. Three equal-weight combinations were tested, with the components chosen by *structural* rules written in advance (every fully-scored US candidate; every batch-3 rule trading exactly the S&P; every fully-scored crypto rule) — never by their scores, because picking components for having scored well is selection wearing a diversification costume. Zero survivors. The best of them is the closest miss of the programme (§5), and the reason combination cannot work with this pool is the independence finding (§3). --- ## 3. The intellectual core: 30 components, 1.8 independent bets **Correlation** measures whether two things move together: +1 means always, 0 means no relationship. A portfolio's whole promise rests on its parts *not* moving together — when one is down, another is up, and the ride smooths out. After the combination verdicts were recorded, every component was simulated standalone and the correlation of daily returns measured between every pair. Across the 30 US components the average pairwise correlation was **+0.55**. There is a standard way to turn that into "how many genuinely different bets is this": N/(1+(N−1)·ρ̄). The answer: > **30 components. 1.8 effective independent bets.** > (The 6-component and 9-component combinations: 1.8 and 1.9.) In plain terms: this ledger contains, in sixty costumes, roughly **two** bets — "be long the stock market, throttled" and "be long crypto, throttled". Adding 24 more components to the US portfolio added *zero* diversification, because a rule that is long the S&P on Mondays and a rule that is long the S&P when the VIX falls are the same bet with different trigger words. This retroactively explains most of the ledger: - **Batch 2's seven MIXED verdicts** were reported as seven near-misses. Three of them "passed" only against XLE, an energy fund that fell 60.6% — and the correlation matrix now shows the four sector-rotation entries (`low-vol-sectors`, `max-effect-sectors`, `idio-vol-sectors`, `52w-high-sectors`) correlate at **+0.94 to +0.98** with each other. Four registry rows, one strategy. The honest count of distinct near-misses in batch 2 was never seven. - `halloween-spy` and `halloween-sectors` correlate at **+0.98** — the same calendar bet aimed at one index versus twelve funds. - `xs-momentum-crypto` and `high-volume-premium-crypto` correlate at **+0.89** — coin momentum and coin volume-spikes pick the same coins at the same times. - And it is why the survivor's sibling failing (§1) mattered so much: **counting costumes as confirmations is the easiest way to talk yourself into an edge.** One measurement caveat, recorded with the result: sleeves that are in the market less than a quarter of the time show *low* correlation mechanically — they are mostly cash — so the true redundancy is, if anything, understated. --- ## 4. Why the null result is credible "Sixty ideas, one survivor" only means something if the measuring instrument could have detected a survivor, and would not invent one. Most write-ups skip this part. It is the part that makes the rest worth reading. **The instrument was attacked before it was trusted.** Adversarial reviewers were pointed at the scoreboard with one instruction: get a worthless strategy to PASS. They succeeded repeatedly, and every route is now closed and pinned by a regression test — **14 documented false-PASS routes**, each of which produced a clean pass on real data before it was fixed. Highlights, because the shapes are instructive: - A strategy that **never traded** passed: a flat account has no drawdown, so its MAR was infinity, which beats any benchmark. - A rule that read **tomorrow's price** passed every gate — there was originally no lookahead defence at all. - **"Hold Solana every Wednesday"** — pure calendar noise — passed, because against a benchmark that *lost* money every ratio-based gate degenerates and the bar collapsed to "make any money at all". The fix: the training window gets a veto, and a losing benchmark triggers an absolute floor instead of a ratio. - Scoring with `record=False` left **no trace**, so forty variants could be tried and one reported as a first-attempt success. Every read of the test window is now logged permanently — a pass on attempt thirty reads as a pass on attempt thirty. **Data lies were tested for with known answers.** Every new series had to contain events that unmistakably happened: the VIX had to spike 6× peak-to-trough in March 2020 (it did: 13.68 → 82.69); the 2-year Treasury yield had to rise ~4 points across 2022 (it did: 0.77 → 4.72); S&P volume had to explode in the March 2020 panic (it did: 6× the 2019 median); Apple's earnings dates had to land on the exact known days (3 of 3 did). Events are measured **peak-to-trough, not endpoint-to-endpoint** — an earlier check measured an oil spike by its endpoints, saw the leftovers instead of the event, and the lesson was pinned. **Two deliberate canaries, both with proven teeth:** - *The junk portfolio.* Averaging many noisy things mechanically smooths the result, so a ratio-based bar could in principle be gamed by combining enough garbage. `combo-junk-calendar` — nine sleeves of pure calendar noise and already-refuted seasonality — was registered with "THIS MUST FAIL" as its success criterion and scored *before* any real combination. It produced a **better Sharpe than the market** (0.97 vs 0.93) and 0.44× the drawdown — dilution really does flatter every risk-adjusted number — and it failed anyway, on the CAGR gate and the 1.30× margin. The method does not manufacture edges from nothing. - *The shifted-VIX canary.* The one failure the lookahead check structurally cannot see is a feed that is off by one day — the check perturbs data *by* date, and the error *is* the date. So `vix-down-day-spy` (hold the S&P the day after the VIX falls — worthless by construction, since VIX changes track the *same day's* market move) was registered as a must-fail. On correct data: +2.6%/yr, MAR 0.10, dead on arrival. On a VIX series deliberately shifted one row early: **+66.4%/yr, MAR 3.24 — an unqualified pass.** A single-row data error would have manufactured the best result in the ledger. The canary failed on the real data, which is evidence the feed is aligned — and the shift test is now a permanent part of the data-layer verification. **A soft spot found later, recorded rather than smoothed over: the near-zero benchmark.** Sibling of the losing-benchmark degeneracy above. The floor that stops a rule "beating" a benchmark that lost money only triggers when the benchmark actually LOST — and against a line that earned almost nothing, every ratio gate degenerates legitimately (11× the CAGR of nearly zero is still nearly zero). The calibration canary for the multi-asset batch exposed it: calendar-parity noise "passed" against holding the euro ETF, which earned +0.3%/yr over the test window, while failing against every real benchmark. This **cannot mint a false PASS** — a PASS requires beating every line, and the hard lines held — but it **can inflate a MIXED**, so any MIXED on a universe containing a flat asset has to be read by asking WHICH lines passed. Lines against near-flat benchmarks are decoration. **The tells behaved.** Momentum and reversal — opposite selections over identical names and dates — did not both pass (both passing is the signature of a harness that deletes losers). All four entries registered as expected-failures failed exactly as registered. **The coin-flip signature, demonstrated by a live candidate.** The cleanest teaching example in the ledger arrived after the core programme: `trend3x-tp50`, a take-profit overlay on the trend3x gate (sell at +50%, re-enter after a full gate reset), registered because a real paper-account trade had given back a $1,000 open profit. On the test half it looks like the best discovery here: **+50.8%/yr at −22.4% drawdown, MAR 2.27**, against plain trend3x's +31.7%/yr at −40.4%. And on the training half the same overlay **loses to the rule it decorates**: +21.8%/yr against +36.1%/yr, MAR 0.58 against 0.64. The two halves of history disagree about whether the overlay helps at all, which is what a coin flip looks like when one side of it happens to land in your test window. The mechanism is mundane: a +50% cap sells a decade-long trend early (the 2010s training years) and shines in a choppy, crash-containing window (the test years). Any process that judged candidates on the test window alone would have shipped it. This is why the training window has a vote, and why a spectacular test number should raise suspicion before excitement — the full detail is in the 2026-08-18 addendum. None of this proves the instrument is perfect. It proves the instrument can detect a positive (the junk and canary results are, mechanically, detected positives on corrupted inputs) and that the known ways to fake one are closed. --- ## 5. The closest miss, honestly `combo-us-all`: the equal-weight combination of all 30 fully-scored US rules, failures included. On its test window (2020-07 to 2026-07): | | CAGR | max drawdown | MAR | Sharpe | |---|---|---|---|---| | the combination | +10.8% | **−11.9%** | **0.91** | **1.14** | | hold the S&P | +16.7% | −24.5% | 0.68 | 0.97 | The best MAR and the best Sharpe anything posted in sixty tests. Less than half the market's drawdown. It passed outright against three *real* benchmarks (developed international, emerging markets, healthcare — not the fell-60% energy fund that flattered batch 2's MIXEDs). And its MAR ratio against the S&P was 1.33× — comfortably above the junk portfolio's 0.89×, so this is **not** the dilution artifact the calibration was built to catch. It is still not a pass, for two reasons that survive any amount of goodwill: 1. **It earned 0.64× the market's return** against a bar of 0.80×. Same epitaph as the other forty-nine: the smoothness is bought by being partially out of the market, and that costs more growth than it saves in pain. 2. **It did not transfer.** On the four sectors none of its components were designed on, it earned +7.0%/yr while just holding them earned +10.3%/yr. Verdict as recorded: MIXED. The honest sentence: *the drawdown diversifies away, and a third of the return goes with it.* --- ## 6. What was not tested, and why **US individual stocks — blocked by survivorship, and worth understanding.** **Survivorship bias**: if your price history only contains companies that still exist, every strategy tested on it looks better than reality, because the bankruptcies are missing from the record. The check was direct: ask the free price sources for eleven famous corpses (Enron, Lehman, WorldCom, SVB, First Republic, Bed Bath & Beyond, …). **Zero of eleven were served.** Nine are simply absent. The other two are the most deceptive thing this project found: **ticker reuse**. Yahoo serves "BBBY" and "SBNY" as healthy series running to today — because after Bed Bath & Beyond and Signature Bank died in 2023, *other companies took over their ticker symbols*. Nothing is wrong with the data as data; it passes every missing-value, gap and staleness check ever written. It is simply a different company's prices under the name you asked for. The only check that catches it is "does the series *end* when the company died" — a question about history, not about data quality. The bitter part: company fundamentals are *fine*. SEC EDGAR serves every number ever filed, with the date it was filed — **point-in-time**, meaning you can know exactly what was public on any given day. The proof is in the verification suite: IBM's Q3-2021 revenue was **$17.618bn as first filed** (2021-11-05) and **$13.251bn as restated** a year later — a −24.8% rewrite (a spin-off reclassification; GE's 2023 revenue moved −48% the same way). A backtest reading today's restated number on the original date is reading the future, and most free fundamental sources serve exactly that. EDGAR doesn't. But clean fundamentals applied to a survivor-only price universe is the same lie with better provenance, so **no fundamental strategy was registered at all** — that was a pre-declared gate, and it held. **Three families remain out of reach:** value/accruals/profitability ranking and post-earnings drift (both need survivorship-free individual-stock prices; not obtainable free), and pre-Fed-announcement drift (tradable ~3% of days — below the 5% exposure floor under which this harness refuses to judge anything, the same floor that made the Santa rally INVALID rather than a verdict). **Vietnam.** Tested honestly and closed early: nothing cleared the bar against buying the HOSE names once and never touching them, after Vietnam's ~1% round-trip costs. The best VN idea (`xs-momentum-vn`, +34.4%/yr at −23.0%, MAR 1.50 standalone) beat several individual names and still failed against the buy-once basket — more return (1.20×) but a deeper drawdown (1.06×) and nowhere near the MAR margin (1.13× against 1.30×) — and its edge did not transfer to the eight HOSE names it was never designed on (+8.8%/yr there vs +21.4%/yr for holding them). An earlier VN trend rule made the same point more bluntly: it beat the monthly-rebalanced benchmark and lost to buy-once-and-drift, +508% vs +930% — the choice of benchmark alone flipped the verdict, which is why buy-once is in every benchmark set here. Two hard local lessons are recorded in the ledger: the only VN index fund's Yahoo feed served **one frozen price through the entire 2022 crash** (a stale print passes every missing-data check — flatness needs its own check), and Vietnam's T+2.5 settlement means shares cannot be sold for ~3 sessions after buying, which mechanically **rejected** a reversion rule that was legal on paper (114 violations). No VN combination exists because only one VN idea was fully scored, and one thing cannot be combined. --- ## 7. What this does not prove Stating the limits as plainly as the findings: - **It does not prove markets are efficient.** Sixty ideas is sixty ideas. The strongest published anomalies in individual stocks were never tested here (§6), intraday behaviour was never tested (daily bars only), and options, futures, bonds-as-an-asset and short-selling were never tested (long-only, cash instruments, no leverage unless declared). - **It does not prove "nothing works".** It proves these 60 rules, on this data, at these costs, under this bar, did not beat buying and holding — with one specific, narrow exception. - **The survivor itself is one draw.** `btc-trend-100d` passed a hard test once, its edge transferred to unsearched coins, and its sibling construction failed — which is exactly what a real-but-specific edge looks like, and also not far from what a lucky draw looks like. It runs live at a fifth of the paper account precisely because the honest confidence level is "worth running small", not "proven". - **The bar is a choice.** 1.30×/0.75×/0.80× are judgement calls fixed in advance. A different bar would pass different things — which is why the junk portfolio and the calendar-noise attacks matter more than the thresholds themselves: they show what this particular bar lets through (nothing, so far) and what it was designed to stop. - Costs are modelled (2bp US, 50bp Vietnam, 10bp crypto per side), not lived. Slippage, taxes, and the experience of actually sitting through a −27% drawdown are not in any number here. --- ## 8. What would make the portfolio question worth asking again The combination attempt failed for a measured reason: the pool contains ~2 independent bets. The fix is not more rules of the same kind — a 51st long-only gate on the S&P would join the +0.55 correlation blob and change nothing. The portfolio question becomes interesting again only when a component would arrive with a **genuinely different return stream**: - **Short volatility** — earning the insurance premium (selling options or their proxies) pays in calm markets and loses in crashes: near-mirror timing to a long-equity gate. - **Duration** — long-term government bonds respond to growth and inflation, not earnings; the 10-year yield data to build the signals is already cached. - **Cross-asset carry** — hold what pays you to hold it, across currencies, commodities and bonds; the classic uncorrelated-return family in the literature. Each of these needs new instruments or new data (options-based funds, bond futures or their ETF proxies, currency pairs), and each would face the same registry, the same bar, the same calibrations. The correlation matrix — not any single verdict — is the test of whether it was worth adding: a real diversifier would drag the effective-bets number off 1.8. That is the number to watch. **The first attempt at exactly this has now been made, and it failed — which is itself a finding.** `gtaa10-faber` (2026-08-18): the canonical published multi-asset trend rule (Faber 2007, verbatim) on a ten-ETF universe declared in advance across equities, Treasuries, gold, commodities, REITs and currencies — built for one purpose, to be a third independent bet next to trend3x and btc-trend, with the bar written into its registry entry before any number existed: effective bets for the three-strategy set had to reach 2.5 of 3. It is the first candidate in this ledger to fail for a **portfolio** reason rather than a returns reason. Its returns did what the literature promises (a quarter of the market's drawdown at a near-market Sharpe; the gates it failed were the familiar CAGR ones). But its correlation with trend3x is +0.49 daily, +0.43 monthly, and the three-strategy set measures **1.97 effective bets of 3** — four of its ten sleeves are gated long equity, and a gated long-equity book moves with a gated leveraged-Nasdaq gate whatever the bonds and currencies do. The rule built specifically to drag the ~2-bets number upward did not move it. The lesson sharpens the paragraph above: breadth alone is not diversification when the breadth is still mostly equity — the missing bet has to come from a different MECHANISM (short volatility, carry, duration), not a wider basket of the same one. Full numbers in the second 2026-08-18 addendum. --- *The equipment survives the programme: `fetch.py --verify` (data integrity, 21 checks), `score.py --selftest` (6 measurement traps + 14 pinned false-PASS routes), `registry.py` (the ledger — 60 entries, failures never deleted, every read of every test window logged). Anyone extending this work should run all three before believing anything, including this document.* --- ## Addendum, 2026-08-18: take-profit and stop-loss overlays on trend3x *Prompted by a real trade: the paper account's TQQQ position was up about $1,000 open profit and gave nearly all of it back while the gate stayed on. The question was whether an exit overlay would have kept that money. Three overlays were registered (`trend3x-tp50`, `trend3x-trail15`, `trend3x-trail15-reentry`), parameters fixed at registration, and scored on the standard rig. All three came back MIXED, and none is shippable. Verdict: the paper account keeps plain trend3x.* The head-to-head on the test window (2021-08-23 to 2026-08-07) looks, at first glance, like a case for the take-profit: sell at +50% and the test half returns +50.8%/yr at -22.4% drawdown, against plain trend3x's +31.7%/yr at -40.4%. But the train half says the opposite: there the same overlay earned +21.8%/yr against plain's +36.1%/yr, with a worse MAR (0.58 vs 0.64). The two halves disagree about whether the overlay helps at all, which is the same coin-flip signature this ledger has already used to fail other ideas. What actually happened is mechanical: a +50% cap costs you most of a decade-long trend (the 2010s train window) and happens to sell near local tops in a choppy, crash-containing window (the test half). Whether the next five years look like the first kind of decade or the second is exactly the thing nobody knows. The trailing 15% stop shows the same disagreement in milder form (train MAR 0.53 vs plain's 0.64, test 0.99 vs 0.78) and gives up a third of the return in both halves. The stop-with-re-entry variant is worse than plain everywhere and worse than its own no-re-entry sibling, because re-entering on a 20-day high buys back into chop repeatedly during 2022. Two further nails, both from the standard gates rather than from judgement. First, none of the three cleared the fixed bar: every one failed the drawdown gate against SPY, and the re-entry variant failed against QQQ too. Second, the holdout: pointed at the four never-searched sectors, all three overlays LOST money (about -3.9%/yr) while holding those sectors made +6.9%/yr, the identical trap-4 failure already recorded against plain trend3x. An overlay cannot fix a rule's transfer problem; it inherits it. The practical answer to the original question: the give-back was not a bug, it is the cost the trend gate pays on purpose, and the overlay that would have kept that one trade's profit would have cost more than it saved over the train decade. The one overlay finding worth remembering if drawdown comfort ever matters more than return: the plain trailing stop did cut the daily test drawdown from -40% to -22% at a price of about 10 points of CAGR. That is a trade you could choose to make with open eyes; it is not an improvement, and it did not pass. --- ## Addendum, 2026-08-18 (2): multi-asset trend following (batch 6) *The first candidate that is not a long equity rule in disguise: Faber's timing model (2007, SSRN 962461, rules quoted verbatim in the registry entry) on a ten-ETF universe declared in advance across equities, government bonds, gold, commodities, REITs and currencies. Registered as `gtaa10-faber` with a calibration canary `gtaa10-junk-months` scored first. Rejected instruments and reasons are written in the entry (USO/UNG for structural roll decay, UUP for double-counting the euro, SLV, AGG/BND, GSG, BDRY). Moskowitz, Ooi and Pedersen (2012) is the futures-native canonical form; its 40%/sigma vol sizing needs leverage and shorting this simulator refuses to price, so the long-or-cash published form was used unmodified rather than inventing a hybrid.* **The calibration came back MIXED rather than FAIL, and the reason is a finding.** Calendar-parity noise on the same universe and sizing beat exactly one benchmark line: holding FXE, which earned +0.3%/yr over the test window. Against a near-zero POSITIVE benchmark every ratio gate collapses (11x CAGR of almost nothing is still almost nothing), and the 0.50 MAR floor that guards against losing benchmarks does not apply because +0.3% is not a loss; the same floor correctly killed the FXY line next to it. Structurally this cannot mint a false PASS, since PASS requires beating every line and the hard benchmarks (SPY, DBC, both basket flavours) all held. But it inflates MIXED: a flat benchmark is a free line, and any MIXED on a multi-asset universe has to be read by asking WHICH lines passed. The noise did not survive the gate, so the machinery is fit to score the real candidate. **The candidate: MIXED, and the shape of the miss matters.** Test window (2020-09 to 2026-08): +5.4%/yr at -6.9% daily drawdown, MAR 0.79, Sharpe 0.91. The published promise (equity-like Sharpe, bond-like drawdown) shows up exactly: drawdown one quarter of SPY's, Sharpe within a tenth of SPY's. What fails is the same gate that killed most of the fifty: CAGR. It earned 0.32x SPY's return against a 0.80x bar, and lost the return race to every risk asset it times (EFA, EEM, GLD, DBC, VNQ). It also failed MAR vs SPY (1.13x against the 1.30x margin). It cleared the bar only against the monthly-rebalanced basket, the two bond ETFs, and the two flat currency trusts. The test window is a near-uninterrupted bull market; Faber's own case is built on 1973-2008, which contained multi-year bears this window did not offer. The holdout does not transfer either: the same rule on the four never-searched sectors made +4.7%/yr while holding them made +9.7%/yr, trap four's exact signature, same as trend3x before it. **The purpose test failed, and that is the headline.** This was built to be a third independent bet next to trend3x and btc-trend. Measured on overlapping days 2017-2026: correlation +0.49 daily (+0.43 monthly) with trend3x, +0.15 (+0.24) with btc-trend. Effective independent bets for the three-strategy set: **1.97 of 3** against the registered bar of 2.5. Four of its ten sleeves are gated long equity, and a gated long-equity book moves with a gated TQQQ whatever the other six sleeves do. It is roughly half a new bet, not a third bet. The registry entry said in advance that high correlation with the existing pair is failure of purpose even on a gate pass, so: failure of purpose. Caveat noted in advance and repeated here: btc-trend is in the market only 53% of days, so its low correlations are partly mechanical. **What this run could not verify.** The ETF window opens in 2007-02 (FXY inception), so the strategy was never scored through a 1970s-style multi-year bear, which is the regime the literature says pays for everything. A verdict on that costs futures data this project does not have. And the correlation was measured on simulated curves over nine overlapping years, not on live records. **Where this leaves the search.** The gap Harry named (an uncorrelated return stream) is still open. Multi-asset trend as implementable in this account is not it: the long-only ETF form keeps enough equity beta to be half the same bet. Streams that could actually be a third bet remain the ones named at the end of the programme: short volatility, carry, duration as its own signal. All are mechanisms other than trend, which the mechanism-balance warning has been asking for anyway. --- ## Addendum, 2026-08-25: add-to-winners pyramiding on SPY momentum (`momentum-pyramid`) *Registered from Tom Hougaard's principle (tradertom.com; "Best Loser Wins", 2022): add to winning positions, never to losing ones. Hougaard publishes the principle, not parameters, so the 1/3 increments, the new-month-end-high add trigger and the 1.0 cap were fixed at registration, before any measurement, and the monthly cadence is a translation of his intraday discretion onto this ledger's TSMOM framework. Verdict: FAIL, and the pre-declared head-to-head against plain TSMOM was not cleared either.* The declared null hypothesis was not buy-and-hold but the plain full-size 12-month rule — the claim under test was the management principle, so the bar was relative: MAR at least 1.10x plain, drawdown no worse, CAGR at least 0.80x. On the test window (2016-07-13 to 2026-08-07) the pyramid earned +11.0%/yr against plain TSMOM's +12.1%/yr with an IDENTICAL -33.7% daily drawdown: MAR 0.91x, DD 1.00x, CAGR 0.91x. The failure mode was named in the registration notes before the run: scaling in gives up early-trend gains, and the 12-month exit still leaves the position at full size at tops, which is exactly where the drawdown lives — so the pyramid pays the entry cost of the sizing and collects none of its protection. The principle's whole case is that failed signals are cheap at 1/3 size, and SPY's monthly signal simply does not whipsaw often enough for that discount to matter. The standard gates add the usual two nails. Against hold-SPY it failed everything (CAGR 0.72x, MAR 0.72x, DD 1.00x), and on the four never-searched sectors it made +3.9%/yr while holding them made +10.0%/yr — the same trap-4 non-transfer already recorded against the trend3x family. A management overlay inherits its base rule's transfer problem; this is now the second demonstration. One caution on what this does and does not refute. It refutes THIS translation — monthly cadence, 1/3 steps, SPY — not Hougaard's intraday practice, which is discretionary and therefore not testable on this rig at all. But the direction of the failure is instructive: the principle is sold on cutting losers cheaply, and at monthly resolution on an index that grinds upward, the cost of arriving late to every trend is larger than everything the small entries save. --- ## Addendum, 2026-08-25 (2): short volatility, the pre-named third stream (`svxy-contango`) *Section 8 named short volatility as a stream that could be a genuinely different bet, and this entry tested it: the Simon & Campasano (2014) contango signal — already FAILED here when pointed at SPY — pointed instead at the instrument the paper actually means, a short-VIX-futures ETF (SVXY, fetched for this test). Signal wording copied verbatim from the failed sibling so the entries differ only in what is traded. Verdict: FAIL, the 57th idea and the 56th non-pass.* The number that kills it is not the drawdown, it is the gate's uselessness: the VIX curve sits in contango 93% of days, so the rule is nearly always long and the signal adds nothing but whipsaw. On the test window (2022-02 to 2026-07) it earned +8.8%/yr at -43.7% against +17.4%/yr at -46.4% for simply holding SVXY — the filter cost half the return and kept essentially all of the pain — and against hold-SPY's 0.62 MAR the strategy's 0.20 is not close. The train half (+38.7%/yr) is carried by the pre-2018 -1x era, an instrument that no longer exists; the entry declared that leverage change in advance and the test half answers for the instrument that does exist. The honest reading for §8: 'short volatility' as implementable in this account — a long-only position in a half-leverage ETF, timed by a signal that is almost always on — is not the diversifying stream the section hoped for. The premium the literature documents lives in futures spreads and options, instruments this programme's account does not trade. That closes the second of the three named directions on the same grounds as multi-asset trend: not falsified as an idea, unreachable in this implementation universe.