Orven

Most quant claims survive a paragraph. Few survive a simulation.

[Orven @orvenn_]

Harvey, Liu & Zhu catalogued 300+ published factors and argued the honest bar under multiple testing is closer to t > 3.0. a large share of the literature sits below it.

and that's conservative — it corrects for what got published, not for what got tested and quietly dropped. the real denominator is invisible.

[Orven @orvenn_]

yesterday: 210 candidates a day, 57 fake alphas a year out of pure noise.

the obvious question is what happens when you scale the swarm. so i ran it.

5,000 candidates a day. same universe, still no alpha in it by construction, same two gates.

1,221 fake alphas a year. 23x the search, 29x the garbage.

but that's not the part that got me.

the winners get better looking as you search harder. median best t-stat climbs from 2.69 to 3.71. at 1,000 candidates a day, 100% of runs produce something "significant." every single day, forever.

meanwhile the honest threshold — the one corrected for how many tests you actually ran — climbs from 3.67 to 4.41.

so the gate stays at 2.5 while the bar it should be clearing runs away from it. the faster your swarm searches, the further behind your validation falls.

forward-year Sharpe of every survivor at every scale: still zero.

scaling the search doesn't find alpha. it manufactures more convincing noise.

code's the same as yesterday's, one parameter changed. run it.

[Orven @orvenn_]

nice, question

yes, and that's the fix nobody ships.

computing it is one line — Šidák gives 3.67 at 210, 4.41 at 5,000. the hard part is the denominator: every discarded variant and hand-tuned parameter counts as a test, and none get logged.

the gate stays at 2.5 because nobody wants to write down how many times they looked.

[Orven @orvenn_]

the correction is Šidák — per-test alpha = 1-(1-0.05)^(1/N). crude, assumes independence, which real candidates aren't. de Prado's Deflated Sharpe does it properly via the variance of trial Sharpes. either way, 2.5 is a single-test bar.

[Orven @orvenn_]

@rektfreedotcom nobody publishes it — dead strategies don't get press releases, so every number you see is a guess.

what I can say: my noise survivors went 1.67 out-of-sample, 0.06 forward. the drop isn't random — it scales with how many tests you ran before the winner showed up.

[Orven @orvenn_]

This article teaches you to build a machine that tests hundreds of factor hypotheses a day.

A man who ran $13 billion recorded a free lecture on why that exact machine kills you.

Marcos López de Prado. Guggenheim, then AQR's first head of machine learning, research fellow at Berkeley Lab. Seven reasons most ML funds fail. Two of them are backtest overfitting and cross-validation leakage — precisely what the pipeline above automates.

his core warning fits in one line: the algorithm will always find a pattern. even when there is none.

so i built the pipeline in a universe where there is none.

pure noise. every candidate a random return series, true mean zero, by construction. no alpha exists in this data. then i ran the article's exact spec on it — 7 agents, 30 variants each, 210 candidates per 24h run, 250 runs. gate 1: t-stat above 2.5. gate 2: out-of-sample Sharpe degradation under 30%.

52,500 candidates. 57 cleared both gates.

five "validated alphas" a month, discovered in nothing.

76% of daily sweeps produced a winner clearing t > 2.5. that threshold was built for a single test. run 210 a day and it stops meaning "this is real" and starts meaning "this one won the lottery."

but here's the part that should actually scare you.

the 57 survivors averaged a 1.41 Sharpe in-sample. out-of-sample they averaged 1.67. the second gate didn't kill them. it promoted them. an out-of-sample check run on 52,500 candidates isn't validation — it's a second lottery, and it selects for the ones that got lucky twice. that's why they look better after it, not worse.

forward year, on data neither gate ever touched: Sharpe 0.06. 47% profitable. a coin flip.

corrected for how many tests the swarm actually runs, the honest bar is t > 3.67 per sweep, and t > 4.90 across a year of them. the article ships 2.5.

none of this is an argument against graphs. the architecture is fine. it's an argument that automating the search without automating the penalty for searching just industrializes fake alpha — and hands you a Sharpe 1.67 out-of-sample chart to feel good about while it does.

code's below. run it yourself, change the numbers, watch the fake alphas multiply as you widen the sweep.

the lecture is free. it's older than half the tools in that article and it will outlive the rest.

https://t.co/rNhVprIDUN

[Roan @RohOnChain]

https://t.co/8HHG7saiut

[Orven @orvenn_]

@Valdemar1x5c Šidák on the family — that's the 3.67 and 4.41 on the chart. crude since it assumes independence; Deflated Sharpe handles correlation properly. either way 2.5 is a single-test bar.

[Orven @orvenn_]

@aislopreality still a test. stops being validation once you select on it.

keep the winners out of 50k and the OOS window is just search space. my noise survivors went 1.41 in-sample → 1.67 out — it picked the double-lucky ones, not the real ones.

[Orven @orvenn_]

@rektfreedotcom good question — bit of both honestly. it lowers the barrier for everyone but most people still won't actually start, so the competition never really shows up. the ones who just do the thing still win easily :)

[Orven @orvenn_]

@19996093ehi yeah this is it. same tools for everyone — the edge is the system you point them at. execution's the whole moat now.

[Orven @orvenn_]

@Denisevuk3 yeah man this is the whole thing right here.

[Orven @orvenn_]

Everyone is looking at the snowboarder. I am looking at the $300 million that just became a prompt.

One line of text. Opus 5. Real sliding physics, zero glitches, first try. In 2012 this exact game was Subway Surfers — 40 people, 12 months, and 3 billion downloads into one of the highest-grossing mobile games ever made.

Today it is a sentence you type before your coffee cools. The part that used to cost a studio a year now costs 30 seconds.

But the game was never the hard part. Building it was the moat, and that moat just dropped to zero for everyone at once. 4 billion phones, and the barrier to fill them just went from $500k to $20.

Most people will watch the snowboard and scroll. A few will do the math on what a 2012 gold rush costs to start in 2026.

[Alex Ermolov @alex_erm]

Opus 5, snowboarder test, one shot. On par with Fable, ahead of every other model I've run. No visual defects on the first pass, and the sliding physics feel right.

[Orven @orvenn_]

Jim Simons, mathematician, former codebreaker, and founder of Renaissance Technologies — the most successful hedge fund in history:

"Trading on gut turns your insides out. A genius one day, a fool the next. So we built a machine instead — and no human is ever allowed to override it."

most people watch this and hear a nice story.

a few catch the part he almost throws away — the actual reason he beat a market no one is supposed to beat.

it's hiding in plain sight. that's why nobody uses it.

everyone hunts the grail in the wrong place. the secret indicator. the insider tip. the perfect call.

the real edge was never hidden. it was just too boring to believe.

an advantage so small it looks like a rounding error — 5% at the roulette wheel, 50.75% at Medallion. then volume, sizing, and capital quietly turn that sliver into a mathematical certainty.

the casino runs the exact same machine. the guy chasing the "genius call" is just fuel it burns.

Simons didn't predict the future. he found the thing everyone was staring at and ignoring — and never touched it again.

that's the grail. and it's already in the video.

The Casino Always Wins, and So Does Wall Street: The Boring Secret Everyone Sees and No One Uses

https://t.co/LTPdfU8rXv

[Voltex @VoltexGar]

https://t.co/mvYfX85y61

[Orven @orvenn_]

Edward O. Thorp, Mathematics Professor and "Godfather of Quants": "If you try to use evidence to draw conclusions and make judgments and think through things rather than following what 'experts' say, you'll make better decisions for the long run."

this free lecture holds the entire "secret alpha" the trading gurus sell, from the mathematician who pioneered the use of the Kelly criterion in hedge funds.

at the board it's simple. retail traders think they need a magic crystal ball, complex indicators, or insider secrets to beat the market . but actually, the world's most profitable quant funds rely on middle-school probability . they don't try to win every trade; they simply exploit a microscopic 51% statistical edge . they perfectly size their bets using the Kelly criterion to avoid ruin , and let the law of large numbers do the heavy lifting over 10,000 trades . that is the whole system, minus the marketing.

the math has been free and sitting in textbooks for decades . what nobody can sell you is the discipline to think statistically and let the math play out instead of quitting early.

https://t.co/mWY6jD54qw

[delost @thedelost]

https://t.co/KwHgvzZIXJ

[Orven @orvenn_]

A 28-YEAR-OLD QUANT IN SINGAPORE JUST LEAKED NVIDIA’S CLASSIFIED "SPARK" PROTOTYPE, AND IT DESTROYS THE OPENAI API BUSINESS MODEL.

This is a standalone edge-compute brick built specifically for local LLM inference. It is currently in closed beta, strictly available to a handful of enterprise financial partners in Singapore.

A rogue dev managed to secure one for $18,500 via a private hardware syndicate.

Pause at 0:03. Look at the I/O ports and the heat dissipation mesh. This isn't a consumer gadget; it is an enterprise-grade inference node that fits in your hand.

That little block runs a fully uncensored 70B parameter model locally. Zero latency. Zero AWS bills. Complete data privacy.
He is currently running proprietary high-frequency trading algorithms through it, processing 40GB of sensitive order-book data daily. If he sent that data to ChatGPT or Claude, it would cost thousands and leak his firm's alpha.

Cloud AI is a retail trap. The real money is moving offline. The moat is your proprietary data, and you cannot protect your moat if you are sending your data to a public API.

Nvidia is about to monopolize the localized edge-AI market. Bookmark this before Nvidia's legal team scrubs the video off the timeline, and follow me for more underground hardware leaks.