Famous trading ideas — anomalies, folk strategies, vendor backtests, things people swear by — each put through the same protocol and stamped twice: is the signal real? and does it survive real execution and scale? This page is the view from above. It aggregates; it doesn't re-judge — every verdict below links back to the study that earned it.
(Regenerate with python tools/make_bench_figures.py — it parses the
ledger, so it's always in sync. A cell shows its studies by number while
they still fit and just counts them when they don't; the
live map is where you click into
one and read what's inside.)
Counting happens in one place. This page describes shape, not totals. The ledger is the only complete list, the map above is redrawn from it, and anything that needs an exact number should be read off one of those two rather than quoted here — a count copied into prose is a count that goes stale the next time a study lands, which is how this page once opened with a total that was hundreds out of date.
Where the bench ends up, as a share of it. Every study carries both stamps except 14, which is pre-registered with its verdict still pending:
| Investable | Fragile | Mirage | |
|---|---|---|---|
| Real | 1% | 8% | 4% |
| Weak | — | 11% | 23% |
| None | — | 1% | 53% |
(Shares of the bench, rounded to the nearest point, so they don't sum to exactly 100.)
Read it the way the colours tell you to:
- About one famous idea in eight is statistically real. The other seven-eighths are weak (Mixed folded in) or plain noise.
- Four in five are mirages once you try to trade them. Costs, capacity, decay, or the discovery that the "edge" was beta all along.
- The green column is a rounding error, and it holds four distinct flavours. The originals — Storm-Shy, All-Weather, Balancing-Act — are risk-managers, not forecasters. Those added by the "green hunt" lot (591–640) are harvestable return engines, and tellingly none is a crystal ball either: Fallen-Angels is a forced-seller risk premium, Currency-Hedged-Carry is a covered-interest-parity rate-differential identity, and Starting-Yield is duration arithmetic (the yield on your ticket ≈ your next decade). Those added by the "plumbing" lot (913–962) are a third kind again — not a return you go and get, but a cost you stop paying: Total-Cost-of-Ownership prices the fee-versus-spread trade-off by holding period, Hidden-Financing backs out what a leveraged wrapper really charges you to borrow, and Which-Gold shows the cheapest wrapper for one identical metal genuinely wins. The one from the measurement lot (963–1012) is a fourth kind: a discipline that pays for itself. Confirmation — wait k days before acting on a signal — cuts whipsaw round trips from 78% to 30% [981], and the study prices both sides of the trade-off instead of netting them, then admits the winning k is only knowable afterwards. You can bank a premium, an identity, a saving, or a discipline; you still can't see the future.
The single most important cell isn't the green one — it's Real × Mirage: effects that are genuinely there in the data and still can't pay you, and there are far more of them than there are greens. That gap between "true" and "tradable" is the bench's whole thesis, measured — and it has now widened twice: first under the green hunt (chase likely-real premia and most still die at the trading desk — borrow fees, roll drag, one-way bounds you can't stand on), then under the measurement lot, where an effect can be real because it is arithmetic and still unbankable — the 1% bitcoin sleeve [1003] is the cleanest case.
A quieter cell worth a look: None × Fragile — gold [69] and bitcoin [70] flunk the claims made for them (inflation hedge, digital haven) yet keep a Fragile stamp as plain diversifiers. The story dies; the asset survives. The measurement lot added three of exactly that shape: mismatched holiday calendars [973], silver as "gold with a beta" [987] and the hunt for cycles in a Fourier spectrum [1000] — a folk claim that fails, sitting on top of a real thing you still have to handle.
We sorted the bench into rough families. The boundaries are judgement calls (is
the 52-week high a chart pattern or a momentum factor? we said factor) — the
shares below are honest, the taxonomy is approximate. This table is the taxonomy:
tools/make_bench_figures.py parses the family
rows below rather than carrying a family map of its own, so a study missing from
here is a study missing from every per-family count.
* "Survived costs" = stamped Investable or Fragile (alive on paper, even if thin). The complement is Mirage.
Four patterns jump out:
- ML & forecasting is the deadest corner of the bench — nothing in it survives. Every model-driven forecaster — Markov pipeline [10], ARIMA+GARCH [12], neural net [39], the Stock-to-Flow model [84], a Random Forest [138] and the 'AI-powered' ETF [139] — produced an in-sample story and an out-of-sample coin flip.
- Calendar effects are the opposite failure mode: among the most real per capita and almost none tradable. The pattern is genuinely in the data; the trade built on it forfeits more than it captures (42, 55) — or dies the moment it's published (67).
- Momentum, trend and carry don't die — they limp. These families collect Fragile stamps, not Mirage ones — mostly alive-but-thin: premia with a century of literature that one tape can't certify and costs nearly erase.
- The measurement lot inverts the whole table: most of it is real, most of it survives costs — and one entry is investable. Every other family asks does this edge exist? and mostly hears no. This one asks what does the choice of estimator, window, benchmark or rebalance date do to the number you publish? — and the answer is almost always "something real", because these effects are properties of the arithmetic, not of the market. The tradability column is the punchline: knowing that rebalance dates are a lottery, that a correlation matrix is mostly noise or that beta has a half-life makes your measurement honest. It does not make you money. Real is cheap here; the costs line is still where it ends.
These aren't opinions — each one fell out of multiple studies independently.
1 · The edge dies at the costs line, not the signal line. Of the signals that clear the statistical bar, all but a handful still failed or barely survived tradability. The overnight drift is real and untradable [01]; intraday reversal is real with a 3.31 bp break-even that lives in the least-liquid names [33]; the turn-of-the-month premium is real at t = 5.1 and a window-only book — even with its cash leg paid the T-bill — still compounds half of buy-and-hold [42]; IBS snap-back is real and gone at the spread [19]. Beat 6 — could you trade it? — is where almost everything dies.
2 · Survivorship doesn't just flatter results — it manufactures and even inverts them. On a survivor panel of large caps, the lottery effect ran backwards (−10.4%/yr, t = −2.5) [53], the idiosyncratic-vol puzzle inverted decisively [54], the 52-week-high premium came out negative [50], the net-issuance hedge flipped because the decade's diluters were the growth winners [64], and asset-growth showed nothing where the literature's premium hides in micro-caps [44]. When we found a positive result on a survivor panel, we capped it as an upper bound [48] — the bias cuts both ways and we say which.
3 · Post-publication decay is the norm, not the exception. The pre-FOMC drift is the most spectacular case on the bench: 3% of sessions carried 11.5% of SPY's entire cumulative return — until Lucca-Moench published it in 2011 and the drift collapsed from +0.24%/day to +0.09% [67]. The size premium never showed at all on a 39-year tradable proxy [45]; turn-of-the-month faded from 13.8 to 4.8 bp/day after 2008 [42]; betting-against-beta decayed 0.70 → 0.28 [43]; textbook pairs stopped paying once everyone copied them [05]; the vendor's dual-momentum edge thinned right after the publication that sells it [40]; oil-predicts-stocks reads exactly zero out of sample [49]. An anomaly's discovery date is the start of its obituary.
4 · Leverage is never free — the "free lunch" is usually the financing bill, in disguise. Betting against beta needs 2.78× leverage to be market-neutral, and realistic financing drags its Sharpe from 0.47 to 0.02 [43]. The retail CFD markup — charged on the whole notional, not the borrowed slice — costs a levered dip-buyer 2.65 pts/yr [30]. The 3× ETF "free amplifier" tripled the drawdown, not the Sharpe (0.90 vs 0.98, −82% trough) [61]; extending duration for term premium lowers the Sharpe [59]; and vol-targeting the carry trade makes its crash worse [27]. When a strategy's appeal is "same return, just levered," the lender has already priced your idea.
5 · What's green on the bench you earn by managing risk, banking a premium, or reading an identity — never by forecasting. For most of the bench's life the green column was pure risk machinery: scale exposure down when markets get loud [16], balance risk across assets and win on Sharpe not return [68], or hold the plain 60/40 [97]. The "green hunt" lot (591–640) — hand-picked because they should be real — finally added three greens you buy for yield, and the lesson survives them intact: a forced-seller risk premium you get paid to absorb [610], a rate-differential identity a currency hedge hands you mechanically [613], and duration arithmetic where the yield on the ticket is the return [625]. None forecasts anything — you collect a premium or read an identity. And the hunt's own body count proves the rule: chase 50 likely-real edges and most still die at the desk — borrow fees [617], roll drag [619], one-way bounds you can't stand on [621], the most famous alpha on Earth already spent [628]. Nothing on this bench forecasts returns and pays. Several things manage risk, harvest a premium, or bank an identity — and those do.
6 · Before you ask whether the edge is real, ask what the number is made of. The measurement lot (963–1012) went looking not for edges but for the choices buried in every backtest — which volatility estimator, which window, which benchmark, which rebalance date, which bootstrap — and found that those choices move the published number more than most of the "anomalies" on this bench ever did. A rebalance date chosen a week apart changes the result [997]; a correlation matrix estimated from a normal sample is mostly sampling noise [1010]; beta has a half-life and the Blume slope everyone uses to "correct" it turns out to measure the signal-to-noise ratio rather than the stability [1005]; a purged-CV fold boundary quietly invents a negative IC out of signal-free data [1001]; and your Sharpe ratio depends on the currency you happen to bank in [995].
Several of them rejected their own pre-registered hypothesis, and those write-ups were kept rather than rewritten — the data can tell 1% from 5% bitcoin and wants 16.5% [1003]; the "most stocks underperform cash" result is absent on a survivor panel [1006]; time does not diversify, but the small-sample bias in measuring that is larger than the effect [1007]; Sortino and Sharpe rank identically (Spearman 1.000) [1009]. The lot's own measurement errors were caught mid-build and are now pinned by tests that fail if the mistake comes back. That is the lesson in one line: the most reliable way to find a signal that isn't there is to measure carelessly, and these hold up precisely because they are properties of the arithmetic. Exactly one of them is investable.
🟩 The ones that made it. The risk-managers came first; the "green hunt" lot (591–640) added harvestable return engines — the first time the bench's green column contained anything you buy for its yield rather than its calm; the "plumbing" lot (913–962) added savings; and the measurement lot (963–1012) added a discipline.
The risk-managers — they predict nothing and win on risk-adjusted terms: 16 · Storm-Shy — Real × Investable. Scale exposure down when realized vol spikes. It survives robust inference, real costs, capacity, a parameter sweep and a third tape — and note what it is: not an alpha, a risk overlay. 68 · All-Weather — Real × Investable. Risk parity earns the best Sharpe of anything we tested (0.92) with a third of equities' drawdown — by predicting nothing and balancing everything. Half the return of stocks, though: the green is risk-adjusted, not absolute. 97 · Balancing-Act — Real × Investable. The plain 60/40 lifts the excess-of-cash Sharpe over 100% stocks (HAC t = 2.3, a bootstrap CI clear of zero) and halves the drawdown — but it forfeits ~2.6 pts/yr of return, leans on the historic bond bull, and the bonds did not cushion 2022. Risk-adjusted, not absolute.
The return engines — premia and identities you can actually bank, and still not a crystal ball: 610 · Fallen-Angels — Real × Investable. Bonds kicked out of investment grade get dumped by forced sellers; catching them (ANGL over HYG) pays +18.3 bp/mo, HAC t = 2.44 — and it's not duration (β to Treasuries ≈ 0). It survives dropping the entire 2020 wave, a single-year jackknife, and a block bootstrap (CI clear of zero), and it clears costs with ~1.4 bp/yr of drag. A genuine forced-seller risk premium. 613 · Currency-Hedged-Carry — Real × Investable. Two funds hold the same Japanese basket; the hedged one (HEWJ) quietly out-earns the unhedged (EWJ) by the whole US–Japan rate gap — +1.2%/yr at HAC t = 2.7 even before the 2022 hiking cycle, pass-through slope ≈ 1.0. A covered-interest-parity mechanical identity, not a signal — no forecasting, just collecting the differential the wrapper hands you for free. 625 · Starting-Yield — Real × Investable. The 10-year yield on your buy ticket predicts your next decade of bond returns at R² = 0.92, slope t = 12.5, identity slope ≈ 1 — pure duration arithmetic (Bogle/Leibowitz), robust across pre/post-1950 and a drop-one-decade jackknife. One entry per decade, ~$0 cost, effectively unlimited capacity. You don't forecast the yield; you read it off the ticket.
The savings — not a return you go and get, but a cost you stop paying: 920 · Total-Cost-of-Ownership — Real × Investable. The cheapest fund is not the one with the lowest fee: expense ratio and spread trade off against each other, and which wins is decided entirely by how long you hold. The crossover is computable in advance, per holding period, and it is the rare bench result you can act on before you place the trade rather than after. 945 · Hidden-Financing — Real × Investable. Back out what a leveraged wrapper actually charges you to borrow, rather than what the factsheet says. The implied rate is recoverable from the fund's own tape, it is materially above the headline, and it is a cost you can decline by financing the leverage yourself. 961 · Which-Gold — Real × Investable. Several wrappers hold one identical metal, so the return difference between them is pure cost with nothing else in it — the cleanest natural experiment on the bench. The cheapest wrapper wins by exactly the fee gap, persistently, with no forecast involved.
The one discipline — the only green on the bench that is about how you act, not what you hold: 981 · The Price of Waiting — Real × Investable. Requiring k consecutive days of agreement before acting cuts whipsaw round trips from 78% to 30% and trades from 7.1 to 1.9 a year, across all 12 tape × signal cells. What makes it green rather than folklore is that the study prices both sides separately instead of netting them: 3,636 sessions spent in cash while the raw signal was already right, worth −216,670 bps forgone, against +253,703 bps avoided by exiting late. Some k beat the unconfirmed rule on Sharpe in 92% of cells by +0.110 — and the study says plainly that the winning k differs in almost every cell, which is what choosing it in hindsight looks like. Every arm carries the same one-day execution lag, so the comparison is about confirmation and not about being late.
🟨 The honest fragiles — Real signals that survive on paper but are thin, decaying, or capacity-starved. Worth knowing; not worth quitting your job for:
| Study | What's real | Why only fragile | |
|---|---|---|---|
| 48 | Groundhog | Month-of-year seasonality, t = 4.1, undecayed | Survivor-panel upper bound; breaks even near ~19 bp |
| 52 | Smoke-Screen | Accruals: cash-backed earnings win, Sharpe 0.64 | Short-side costs; documented post-2000 fade |
| 56 | Tide-Table | CAPE forecasts 10-year returns (R² 0.28) | A tide table, not a stopwatch — useless at 1 year |
| 59 | Downhill | Term premium, +2.2%/yr over cash | Sharpe 0.32 vs cash's 1.82; 2022 took −23% |
| 63 | Free-Fall | Short-vol carry, +12%/yr (SVXY) | Skew −4.8, one −83% day; five crash days wiped 95% |
| 66 | Inverted | Curve inversion → +1% next 18m vs +16% normal | ~5% of months, a year of melt-up first — no sell button |
| 67 | Fed-Drift | Pre-FOMC drift carried 11.5% of SPY's return | Publication killed it: +0.24%/day → +0.09% after 2011 |
| 71 | Ambush | Confluence of four dead-net edges: +19.6 bp/day at K≥3 (HAC t = 3.1), undecayed, costs defeated by rarity | ~15 trades/yr → +1.2%/yr excess; OOS Sharpe +0.28 under the frozen 0.30 bar |
| 75 | Knee-Jerk | Connors RSI(2) oversold bounce: pooled HAC t = 10.7, beats a coin by +57 bp/trade | Decayed 35% since the 2008 book; long-only beta in a bull market; the 200-SMA filter hurts |
| 103 | Turtle-Trader | The Turtles' Donchian breakout: a real long-side trend premium, HAC t = +11 | Shorts are a structural trap on up-drifting markets; edge ~halved post-publication; years-long drawdowns |
| 106 | Supertrend | The ATR(10,3) daily flip beats a coin (HAC t = +3.3, bootstrap CI clear of zero) | Only the canonical multiplier works (2 and 4 are noise); ~2.5%/yr gross at ~8 flips/yr |
| 110 | Faber-Timing | The 200-day timing rule lifts SPY's Sharpe 0.55→0.73 and halves the drawdown (−55%→−22%) | Pure risk reduction, not alpha; lags in bull markets; whipsaw + switching/tax drag |
| 144 | Permanent-Portfolio | Browne's 25/25/25/25 stocks/long-bonds/gold/cash: a genuinely low-drawdown all-weather mix | Forfeits much of equities' return; leans on the gold + bond bull; risk reduction, not alpha |
| 151 | Stocks-For-Long-Run | The real equity premium holds in every long rolling window (Siegel) | The unit of time is the decade; 20-yr windows can still trail bonds; useless as a timer |
| 173 | Four-Percent-Rule | Bengen's 4% withdrawal survived every historical 30-yr US retirement cohort | Sequence-of-returns risk; today's valuations/yields; non-US history has failed it (Pfau) |
| 203 | Golden-Butterfly | Tyler's 20x5 mix beats SPY on Sharpe (0.68 vs 0.53) with a third of the drawdown | Loses to its simpler parent (Permanent Portfolio); forfeits ~3 pts/yr of return; risk reduction, not alpha |
| 209 | ETH-BTC-Ratio | The 20-day ETH/BTC momentum rotation beats 50/50 (HAC t=+3.7, alpha t=+9.4) | Strapped to a single crypto cycle, -68% drawdown; one regime, not a law |
| 210 | Crypto-Trend | A 200-day timing rule cuts Bitcoin's -83% crash to -70% and beats buy-and-hold on Sharpe (0.90 vs 0.63) | Whipsaws; ~7 switches/yr; ~12-year history; a drawdown shield, not alpha |
| 223 | Same-Month Seasonality | Same-calendar-month return persists: top-bottom decile spread HAC t = +5.57 over 330 months, bootstrap Sharpe CI clear of zero | Survivor-inflated on ~8-stock deciles; ~78% monthly turnover; short leg hard-to-borrow; a live Russell-1000 build dilutes the gross edge |
| 301 | Triple-RSI | The viral "90% win-rate" RSI(5) bounce is genuinely real: SPY +132.8 bp/trade, HAC t = +5.07, survives an honest next-open fill and post-2010, beats a coin by +85 bp | ~3.5 trades/yr, ~7% of the time in the market → ~+4.7%/yr vs the index's +10.8%; the 90% win-rate is the exit's shape, not the edge (a coin clears 62% at a loss) |
| 302 | Lithium-Boom | A 200-day trend overlay on lithium (LIT) is real — net HAC t = +2.09, and trend-timing beats same-exposure random timing at the 100th percentile | Net excess-Sharpe ~0.51 (bootstrap CI [0.02, 0.99] barely clears zero), trails SPY buy-and-hold; the only real benefit is a halved drawdown (−66%→−41%) — crash-dodging, not skill |
| 340 | Bank-Loans | Floating-rate loans (BKLN) really do dodge duration: beta to long Treasuries = −0.055 (HAC t = −2.95), and BKLN gained +13.9% through the 2020–23 bond repricing | The risk just moved duration→credit (equity beta +0.20, t = +4.48); a thin sleeve (CAGR 3.7%) that fell in 7/7 equity crashes and gapped −24% in March 2020 |
| 363 | PEAD-Drift | Post-earnings drift sorted on the EPS surprise: +1.34% at 20 days (t = 2.96), survives a quarter-block placebo | Nothing in week 1; net edge only at 20–60-day holds on a 30-name survivor basket; thin and long-leg-dependent |
| 367 | CEF-Discount | Widest-discount closed-end funds beat the narrowest by +6.6%/yr, market-neutral (Welch t = 3.72), placebo-clean | A NAV proxy on a survivor basket; fades to t = 1.56 post-2010; hard-to-borrow shorts, ~18 tiny funds |
| 375 | VXX-Roll-Decay | Shorting a VIX-futures ETP's contango is a real carry: +0.171%/day, HAC t = 2.51, +35%/yr net | Skew −1.65, a −43% day and a −92% drawdown make a constant-notional short un-allocatable |
| 601 | Factor-ETF-Live-Test | The factor ETFs do deliver their exposure live: USMV cuts vol to 0.80× (t vs 1 = −7.5), MTUM/VLUE/QUAL load their factor at HAC t = +7.5/+10.0/+8.8 | Exposure is real, alpha isn't — none reliably beats SPY once you pay for it; a delivered risk profile, not free return |
| 618 | GBTC-Premium-Cycle | All three regimes verified to the digit — +36% premium (t = 9.0), −24% discount (t = −7.4), then par; the 2023 discount→par convergence was a real, dated trade | A one-off wrapper lifecycle that ended at the Jan-2024 ETF conversion; not repeatable now |
| 622 | Thematic-ETF-Curse | A calendar-time book of 48 thematic ETFs loses −16.3%/yr CAPM alpha (HAC t = −3.27) in their first 36 months; the broad-index-launch placebo is clean | It's a short-the-hype signal: costly-to-borrow small names, clustered 2020–21 launches, survivorship flatters the long-only escape |
| 626 | Unemployment-Trend-Timing | Gating Faber's 200-day rule on rising unemployment beats the pure rule by +12.3 bp/mo (HAC t = +2.32) and cuts whipsaw spells 64% over 928 months | Most of the edge is trading less, not seeing more; current-vintage unemployment (revisions unmodeled); a smoother Faber, not alpha |
| 628 | Buffett's-Alpha | The most famous alpha on Earth replicates: +9.5%/yr CAPM alpha, HAC t = 3.47 over 46 years (the FKP t > 3 holds to 2011) | The fade is itself significant (−11 pp/yr post-2010, t = −2.38) — quality + low-beta + cheap leverage, mostly spent; little left for today's buyer |
| 632 | Crypto-XS-Momentum | Last week's crypto winners keep winning: +164 bp/wk on the winners-minus-losers quintile, HAC t = 3.56 over 449 weeks on a 44-coin panel including dead pairs (LUNA shorted) | Rides the 2017/2021 bulls; ~55% hit rate drowned by 10–50 bp crypto spreads; paid roughly nothing through the LUNA/FTX year |
| 806 | Prospect-Theory Value | The Barberis-Mukherjee-Wang cumulative-prospect-theory value predicts returns negatively: long low-TK / short high-TK earns +138.7 bp/mo (HAC t = +3.15 over 137 months, positive in both eras, correct BMW sign) | Survivorship flatters the short (blown-up lottery names absent) and a 50-name universe concentrates it into hard-to-borrow lottery mega-caps; +124 bp/mo net at 5 bps but real borrow/squeeze exceeds the charge |
| 812 | Corwin-Schultz Spread | The high-low bid-ask illiquidity estimator earns a premium — high-spread names out-earn by +4.45 bp/day (HAC t = +3.24, both eras, +4.88σ placebo) | The illiquid long leg pays the very spread it earns: net +2.31 bp/day at 1 bp (t = +1.59, insignificant), −5.69 at 5 bps |
| 861 | Debt-Maturity Rollover | Firms funded with a high share of SHORT-TERM debt under-earn the long-funded — low-minus-high-rollover-share tercile +45 bps/mo (NW t = +3.22), right-signed, both regimes (post-2017 +2.83, post-2019 +2.51), ~2.6× stronger in the 2022+ hiking era | Thin ~26-name survivor cross-section (no delisted rollover-wall casualties), monthly turnover + hard-to-borrow shorts; a real balance-sheet risk premium too small and capacity-starved to bank cleanly |
| 888 | CLO AAA Carry | The senior AAA CLO tranche (JAAA) pays a genuine low-vol carry over cash — +1.38%/yr on 1.63% vol, excess Sharpe +0.84 (HAC t +2.33, block-bootstrap CI clear of zero), topping the excess-Sharpe race and beating the un-tranched loans it's carved from | ~5.7y stress-free sample (misses the 2020 CLO mark-down), and all the carry lives in the high-rate era (ZIRP excess ~0) — mechanically a short-rate+spread, so regime-bound and short-history; thin and not yet crisis-tested |
| 889 | Broad Dollar-Hedge Overlay | Generalising 613: on broad developed-international (HEFA/DBEF vs EFA, the same basket) the hedged-minus-unhedged return IS the US–EAFE rate differential mechanically (β on −fx ≈ 1, high R², HAC t clears the bar, bootstrap CI clear of zero) | The identity is robust but the whole tape sits in one dollar-regime (US out-yielded EAFE); the tradable overlay is regime-limited and the pickup is small — a mechanical identity, not a forecast |
This page will be wrong eventually — that's the design. Every verdict on the bench is a falsifiable claim (bar 14, still pre-registered), each with reproducible code, pinned data fingerprints, and the exact line where we think the dream dies.
- Think a Mirage is tradable? Fork the study, change the cost model or the venue, and show the break-even. Beat 7 of every notebook says what we'd consider convincing.
- Think a None is real? The inference stack (HAC, Lo, bootstrap, Reality
Check) is in
quantlab/— run it on your variant. - Got a candidate for the queue? Open an issue. The ideas that look most embarrassing to test are usually the best ones.
The map gets a new chip every time. python tools/make_bench_figures.py
redraws it.
Part of Open-Alpha-Lab. Counts generated from the
ledger by tools/make_bench_figures.py,
and checked against the studies themselves by
tools/check_reference_table.py.
Not investment advice — research and education. See LICENSE.