Here is a shorter version of Claude's take on this thread:

1. Four years are the strategy. From lizgdal's table (5611): 1991, 1998, 1999 and 2020 compound to 32.8x by themselves. Back them out of a 41-year record at ~23% and the other 37 years run at 13-15%, versus ~11-12% for the S&P. So a 2-3 point gross edge plus four melt-ups — two of which are immediately followed by -40% and -25% years. The tail that makes the CAGR is the tail that makes the -66% drawdown.

2. The error bars swallow every parameter choice. At ~30-35% vol over 40 years, SE on CAGR is ~4.7 points. The entire 22.9%-to-27.4% spread across variants fits inside one SE. On Sharpe, SE is ~0.20; after a few hundred configurations the best-of-noise sits ~0.65 above truth, so an observed 1.06 is consistent with a true 0.4 — under SPY's. Count the knobs: lookback, N, HTD, timing index, SMA length, sell band, signal frequency, trade-day offset, three ADV filters, stops, a second ranking metric, four ways to combine screens. And the timing rule is fit to 3-5 bear markets in 41 years. rayvt says exactly that, then picks a winner from the grid anyway.

3. Three results that can't be right. Day-of-month at 27.61% vs 21.24% means ~50bp/month free from moving execution one day — real effects are single-digit bp. rayvt found the overlap backtest leaked (gprc(1)→(2) moved CAGR 250bp, the next lag day ~10bp) and said so in 5698; the 32.2% from those links was quoted again in 5799. And the same strategy, same period, same rule appears as 19.8%, 21.75% and 27.94% in three posts. Meanwhile the reported numbers climb monotonically, 25.6→32.2→38→45.4→48%, over five weeks on one dataset. Discovery doesn't do that. Searching does.

4. Costs and taxes are wrong where it matters. 0.1-0.25% friction doesn't survive pre-decimalization Nasdaq spreads — five names, full monthly turnover, realistically 10-20 points a year pre-2001, not 1-3. That's where the outperformance lives, and it's most of the "decay since 1999." On tax, -4.8 points on 25.6% implies ~19% effective on short-term gains; 8-10 is realistic, which eats the rest in a taxable account.

5. Wrong benchmark, and the "holy grail" line. A 35%-vol strategy shouldn't be measured against unlevered SPY; the honest hurdle is ~2x QQQ with the same filter. "3x SPY with lower beta" is a category error — low beta at double the standard deviation means the risk is idiosyncratic, five names deep in one industry cycle. And "in backtests you can get 3x SPY... as close to the holy grail as you're going to get" — the caveat is the whole story. Something that only exists in backtests is a fit. Same for "everything improved simultaneously," which is the diagnostic for overfitting. Real improvements trade something off.

6. Nobody answered the arbitrage question. rayvt asked it in 5503 and moved on. There's a good answer — no institution holds 5 names at 35% vol through 30-month cash stretches — but it reframes the result as payment for risk nobody else will take, not an inefficiency, and caps the honest excess at about the 2-3 points from section 1.

7. Two research processes here, not one. One leaves artifacts: stated formulas, GTR1 links others can break, explicit friction and periods, and self-corrections — errors found and reported, good results talked down rather than up. The other leaves nothing checkable: no code, no spec, errors fixed conversationally mid-stream, and "hundreds of hours of conversations with it" describes a long search reporting its best draw each time with the rejects unrecorded. Not about effort or good faith — about what output can support. For each number in this thread, ask whether you could reproduce it from what's posted. If not, treat it as something to check, not a finding.

What would settle it: fit every parameter on 1985-2005, apply it mechanically to 2006-2026, report whatever comes out. rayvt proposed this in 5796 and set it aside. Still the one test nobody has run.

What survives: momentum on a large-cap universe with a long-MA filter, worth a few points a year gross at double index vol, -45% to -65% drawdowns, multi-year stretches in T-bills. Defensible at 10-25%, in a Roth. Above that, you're looking at the researcher's fingerprint, not the market's.

Worth adding: the most useful thing in 140 posts is rayvt's rolling-window tables (5519, 5602). That's the right way to present a strategy like this — every start date, not one cherry-picked run — and what they show is a good deal less exciting than the headlines.