Backtest Forensics · @droskill
Someone makes a striking claim about the market — a chart, a thread, a newsletter. I rebuild it from clean, survivorship-free data and test what actually holds up. Usually that means crediting the real insight first, then following it one step further than the original did. Backtests in RealTest and Python; data from Norgate.
14 studies
Sep 2026 · factor investing
When aggressive, high-beta stocks are beating defensive, low-vol ones, that’s supposed to be “risk-on.” Rebuilt from S&P 500 stocks back to 1994: a real 32-year pattern that lives almost entirely in bear-market tails — and a coin flip the moment I tried to actually trade it.
real crash signature, no tradable edgeRead →Sep 2026 · technical analysis
The Industrials and the Transports have to confirm each other — the oldest trading rule there is, and one I’d never seen backtested. Coded as mechanical rules and run over 126 years, it beats random timing and cuts the worst drawdown from −87% to −54%, but trails buy & hold since 1953, and the Transports carry most of the signal.
a crash-avoider, not a return-boosterRead →Sep 2026 · market breadth
The “breadth is thin” posts show the breadth and skip the part that matters: what happened next. A year of tests, gathered in one place — after the scariest thin-breadth warnings the S&P 500 went on to normal-to-better returns, because thin breadth is the fingerprint of a giant-led bull, not a sell signal.
a fingerprint, not a sell signalRead →Sep 2026 · housing & markets
“Housing drives the economy” is repeated like a law — and it’s half right. But across 40 years, stock bull markets barely need a housing boom (the 1990s and 2010s prove it), most bear markets sail right past housing, and the “rates sink home prices” link is mostly a mirage.
busts drive recessions, not bullsRead →Sep 2026 · education & economics
It feels obvious that richer states should read and count better. Across all 50 states + D.C. on the 2024 NAEP, income is a loose predictor (r ≈ 0.3–0.4) and per-pupil spending is basically flat — what a state does with the money beats how much it has.
a little; spending, not at allRead →Sep 2026 · rates & bonds
The pitch has three legs — yields must mean-revert, the record short will squeeze, foreigners are fleeing. Each has a real instinct, but chased down they point to carry and recession insurance, not a coiled spring — and to cash over long duration in today’s regime.
carry & insurance, not reversionRead →Sep 2026 · market breadth
The famous crash omen fires when the market makes new highs and new lows at once. The breadth-divergence instinct is real — but on 31 years of data it’s a mild one-month chill, and as a crash bell it’s an 81% false alarm even at its worst.
a yellow light, not a sirenRead →Sep 2026 · market breadth
A note argues that when the net-new-highs trend is rising, returns run about three times higher. The gap is real — but coincident. Lag the signal one day, so you could actually trade it, and it disappears.
coincident, not predictiveRead →Sep 2026 · market breadth
A table says new highs are the thinnest share of S&P 500 market cap at any peak since 1973. Rebuilt from point-in-time data, “thinnest ever” depends on how you weight it — and thin-breadth highs have paid off just fine.
it depends how you weight itRead →Sep 2026 · gold & money
A widely-shared thread says gold is a record 171% of the money supply, and that every prior spike preceded a monetary reset. The ratio isn’t at a record — and historically it has marked gold’s tops, not the world’s.
not a record, marks topsRead →Sep 2026 · trend following
Price-above-the-average and the average’s slope are two different trend signals. Run Faber-style on a diversified basket since 1988, which one you pick is a dial between staying invested longer and cutting risk earlier.
a dial, not a winnerRead →Sep 2026 · technical analysis
Stan Weinstein’s four-stage framework is a classic. I mechanized it and tested it on 900,000 survivorship-free stock-months to ask a blunt question: how much does the stage taxonomy add over a plain 30-week moving average?
mostly the moving averageRead →Aug 2026 · backtest forensics
A survivorship-safe re-run of an X thread’s 52-week-high breakout study across the S&P 500 and Nasdaq 100. The ATR-breakout edge is real in direction — but roughly half the size once the delisted names are put back.
real, but smallerRead →Aug 2026 · seasonality
A post noted that the best Septembers cluster above the 200-day average. It’s a real filter — but run across all twelve months, it lifts every one of them. It’s a trend effect wearing a seasonal costume.
trend, not seasonRead →A live market dashboard — relative strength and breadth, regenerated nightly.