Olivia
Online
MindArch
@mindarchx
Pure alpha knowledge
State of Mind
Joined July 2026
26 Following
69 Followers
20 Posts
Joe Blitzstein, Harvard professor and author of the probability textbook most quant desks quietly learned from:
"Ten simple yes-or-no choices in a row produce over a thousand possible outcomes.
Quant desks running strategies with dozens of parameters are sitting on numbers so large no one has actually checked most of the scenarios their own model can produce."
this is the exact math behind why "we stress tested every scenario" is almost never literally true.
cut through the notation and the growth rate is brutal. each additional binary choice doesn't add outcomes, it multiplies them. ten choices already produces over a thousand combinations, and that number explodes exponentially from there.
a model with a modest number of parameters can generate more scenarios than anyone could enumerate in a lifetime, let alone one afternoon before a launch.
zoom out to what "comprehensive backtesting" actually means at that scale. testing a sample of scenarios and testing every scenario are two completely different claims, and the gap between them grows exponentially larger with every parameter added to a strategy.
most people reading "extensively tested across scenarios" assume something close to exhaustive coverage, when the combinatorics guarantee that's essentially impossible past a small number of variables.
this is exactly what risk reports skip. the parameter count sounds thorough. the actual fraction of the scenario space that was ever touched is the number nobody puts in the writeup, because it's usually a rounding error.
more parameters sounds like more rigor. it mostly just means fewer of the possible outcomes were ever checked.
Bookmark this legendary professor
https://t.co/ZGpOvAef8u
Nassim Taleb, former derivatives trader and NYU professor of risk engineering:
"The best predictor of a fund's blowup isn't volatile returns. It's steady returns. I traded derivatives for twenty years and watched the smoothest equity curves turn out to be the ones hiding the most risk, not the least."
this is the exact signal every "look how consistent our Sharpe ratio is" pitch quietly gets backwards.
cut through the marketing and the finding is uncomfortable. real markets aren't stable. a fund posting suspiciously flat monthly returns usually isn't operating in a calm environment, it's paying to make an unstable one look calm, and that hidden cost eventually comes due in one lump sum.
the volatility never actually disappeared. it got warehoused, usually through leverage or an unhedged tail position nobody was stress testing, right up until the month it detonates all at once.
zoom out to how allocators actually evaluate track records today. a smooth, low-variance equity curve reads as the safest pitch a fund can make to an LP. taleb's own point is the opposite: unnatural smoothness is one of the strongest tells that a fund is short volatility somewhere off the reported chart.
most allocators looking at a clean, steady performance line assume smoothness means safety, when the smoothing itself is usually the exact thing worth underwriting before wiring capital.
this is exactly what due diligence decks skip. "consistent, low-volatility returns" is the entire pitch on slide one. it's frequently the cost of concealing the one tail risk that will eventually show up all at once, and cost investors everything in a single month.
the flat curve looked like proof of skill. it was often just proof that the risk hadn't been marked yet.
Bookmark this alpha
Nassim Taleb, former derivatives trader and NYU professor of risk engineering:
"The best predictor of a fund's blowup isn't volatile returns. It's steady returns. I traded derivatives for twenty years and watched the smoothest equity curves turn out to be the ones hiding the most risk, not the least."
this is the exact signal every "look how consistent our Sharpe ratio is" pitch quietly gets backwards.
cut through the marketing and the finding is uncomfortable. real markets aren't stable. a fund posting suspiciously flat monthly returns usually isn't operating in a calm environment, it's paying to make an unstable one look calm, and that hidden cost eventually comes due in one lump sum.
the volatility never actually disappeared. it got warehoused, usually through leverage or an unhedged tail position nobody was stress testing, right up until the month it detonates all at once.
zoom out to how allocators actually evaluate track records today. a smooth, low-variance equity curve reads as the safest pitch a fund can make to an LP. taleb's own point is the opposite: unnatural smoothness is one of the strongest tells that a fund is short volatility somewhere off the reported chart.
most allocators looking at a clean, steady performance line assume smoothness means safety, when the smoothing itself is usually the exact thing worth underwriting before wiring capital.
this is exactly what due diligence decks skip. "consistent, low-volatility returns" is the entire pitch on slide one. it's frequently the cost of concealing the one tail risk that will eventually show up all at once, and cost investors everything in a single month.
the flat curve looked like proof of skill. it was often just proof that the risk hadn't been marked yet.
Bookmark this alpha
https://t.co/ZGpOvAef8u
John Geanakoplos, Yale professor who ran a mortgage hedge fund through the 2008 crisis:
"Quant funds size positions like the biggest player sets the price. I watched that assumption fail during the mortgage crisis. The price was never set by the biggest buyer. It was set by whoever was standing right on the edge, barely willing to trade at all."
this is the exact pricing mechanism most position-sizing models quietly get backwards.
cut through the notation and the idea is almost counterintuitive. the price of anything isn't an average or a function of the largest order. it's set by the marginal participant, the trader right on the edge between wanting in and staying out.
make the biggest buyer in the market want it even more, and the price doesn't move. they were never the one setting it.
zoom out to how this plays into execution risk on a live desk. a fund can build enormous size and still have close to zero price impact, because size alone was never the mechanism. only being the marginal trade actually moves the number.
most people watching a large order hit the tape assume the size is what's moving price, when the real driver is happening entirely at the margin, invisible to anyone just counting notional.
this is exactly what "we're the biggest player so we set the price" assumptions get wrong. being big is not the same claim as being marginal, and conflating them is where execution models quietly mis-forecast slippage.
size gets the attention on the trade blotter. the marginal order was always the only one that actually mattered.
Bookmark this.
Ian Ball, MIT game theory professor:
"Every trading algorithm claims it has a strategy. Most of them only specify what to do in the scenarios someone actually thought to test. A real strategy has to define what happens at every single contingency, including the one nobody remembered to code for."
this is the exact gap between a backtest and a strategy that blows up the first time markets do something new.
strip away the jargon and the definition is almost uncomfortably strict. a strategy isn't a rule for the situations you expect. it's a complete plan covering every single contingency you could possibly land in, even the ones you never anticipated.
miss even one of those contingencies and you don't have a strategy at all. you have a rule that happens to work until it runs into the gap nobody filled in.
zoom out to why this matters for anyone running capital. two decision points with two choices each already produces eight full contingency plans for one player alone. add more branching market conditions and the count explodes exponentially almost immediately.
most people staring at a clean backtest curve assume the strategy behind it accounted for everything, when most of the time it only ever specified what to do in the paths that actually got tested.
this is exactly what "our system has a defined strategy" pitches skip. having a rule for the expected cases is not the same claim as having a complete contingent plan for every node in the tree, and that gap is precisely where live trading diverges from the backtest.
a rule looks like a strategy right up until markets find the contingency you forgot to write down.
Save this gem for later.
Ian Ball, MIT game theory professor:
"Two quant desks can see the exact same order flow data and correctly take opposite positions. Neither one is wrong. The entire disagreement comes down to a probability estimate neither desk actually discloses, and that number is the real trade, not the position itself."
this is the exact gap between "the model says buy" and the actual reasoning that produced it.
strip away the dashboards and the math is uncomfortable. once you assign a probability to how the other side of the market behaves, one position becomes correct. shift that probability by a few points and the opposite position becomes correct instead, using the identical data.
there's no position that's right independent of a belief. the ranking of trades pivots entirely on a number that gets assigned before any trade is placed, not on some fixed fact sitting in the data.
zoom out to how often two systematic strategies diverge on the same signal. it's rarely because one has better data. it's because each embedded a slightly different probability estimate into the same information, and that hidden number did all the actual work.
most people defending a trade point to the signal itself, without ever stating the probability threshold that made that particular signal decisive.
this is exactly what confident trade write-ups skip. the position gets justified. the belief that made it optimal, and the exact point where a slightly different belief would have flipped it, almost never gets written down.
the trade was never fixed by the data alone. it was only ever correct relative to a number nobody disclosed.
Bookmark this gem.
@QuantSanchez indeed, i genuinely respect Ian's perspective on this one
Ian Ball, MIT game theory professor:
"Two quant desks can see the exact same order flow data and correctly take opposite positions. Neither one is wrong. The entire disagreement comes down to a probability estimate neither desk actually discloses, and that number is the real trade, not the position itself."
this is the exact gap between "the model says buy" and the actual reasoning that produced it.
strip away the dashboards and the math is uncomfortable. once you assign a probability to how the other side of the market behaves, one position becomes correct. shift that probability by a few points and the opposite position becomes correct instead, using the identical data.
there's no position that's right independent of a belief. the ranking of trades pivots entirely on a number that gets assigned before any trade is placed, not on some fixed fact sitting in the data.
zoom out to how often two systematic strategies diverge on the same signal. it's rarely because one has better data. it's because each embedded a slightly different probability estimate into the same information, and that hidden number did all the actual work.
most people defending a trade point to the signal itself, without ever stating the probability threshold that made that particular signal decisive.
this is exactly what confident trade write-ups skip. the position gets justified. the belief that made it optimal, and the exact point where a slightly different belief would have flipped it, almost never gets written down.
the trade was never fixed by the data alone. it was only ever correct relative to a number nobody disclosed.
Bookmark this gem.
Ian Ball, MIT game theory professor:
"Every trading algorithm claims it has a strategy. Most of them only specify what to do in the scenarios someone actually thought to test. A real strategy has to define what happens at every single contingency, including the one nobody remembered to code for."
this is the exact gap between a backtest and a strategy that blows up the first time markets do something new.
strip away the jargon and the definition is almost uncomfortably strict. a strategy isn't a rule for the situations you expect. it's a complete plan covering every single contingency you could possibly land in, even the ones you never anticipated.
miss even one of those contingencies and you don't have a strategy at all. you have a rule that happens to work until it runs into the gap nobody filled in.
zoom out to why this matters for anyone running capital. two decision points with two choices each already produces eight full contingency plans for one player alone. add more branching market conditions and the count explodes exponentially almost immediately.
most people staring at a clean backtest curve assume the strategy behind it accounted for everything, when most of the time it only ever specified what to do in the paths that actually got tested.
this is exactly what "our system has a defined strategy" pitches skip. having a rule for the expected cases is not the same claim as having a complete contingent plan for every node in the tree, and that gap is precisely where live trading diverges from the backtest.
a rule looks like a strategy right up until markets find the contingency you forgot to write down.
Save this gem for later.
@Rossst_03 alpha as always, good one
@Rossst_03 ty man, Ian Ball know's about game theory a lot
Ian Ball, MIT game theory professor:
"Every trading algorithm claims it has a strategy. Most of them only specify what to do in the scenarios someone actually thought to test. A real strategy has to define what happens at every single contingency, including the one nobody remembered to code for."
this is the exact gap between a backtest and a strategy that blows up the first time markets do something new.
strip away the jargon and the definition is almost uncomfortably strict. a strategy isn't a rule for the situations you expect. it's a complete plan covering every single contingency you could possibly land in, even the ones you never anticipated.
miss even one of those contingencies and you don't have a strategy at all. you have a rule that happens to work until it runs into the gap nobody filled in.
zoom out to why this matters for anyone running capital. two decision points with two choices each already produces eight full contingency plans for one player alone. add more branching market conditions and the count explodes exponentially almost immediately.
most people staring at a clean backtest curve assume the strategy behind it accounted for everything, when most of the time it only ever specified what to do in the paths that actually got tested.
this is exactly what "our system has a defined strategy" pitches skip. having a rule for the expected cases is not the same claim as having a complete contingent plan for every node in the tree, and that gap is precisely where live trading diverges from the backtest.
a rule looks like a strategy right up until markets find the contingency you forgot to write down.
Save this gem for later.
https://t.co/ZGpOvAef8u
Peter Kempthorne, MIT professor whose volatility modeling course trains the quants who price every option on Wall Street:
"Model a stock price directly and every desk gets it wrong the same way. Model the log of the price instead, and the exact same math suddenly works. I ask every new student why, and most can't answer on day one."
this is the one modeling choice separating a pricing model that works from one that quietly fails.
strip away the notation and the reason is almost embarrassingly simple. a ten dollar stock and a five hundred dollar stock don't move by the same number of dollars on an average day, but they do move by roughly the same percentage.
model the raw price and you're implicitly assuming both stocks share identical absolute swings, which is false the moment you check real data.
play that forward to any risk system running today. this single substitution, log price instead of raw price, is the quiet foundation under nearly every option pricing model still in use.
most people staring at a price chart never ask which scale the model underneath it was actually built on.
nobody markets "we picked the correct scale for our variable" as an innovation. it sounds like a footnote, even though picking the wrong one silently breaks every assumption built on top of it.
The model gets the credit when it prices correctly. The scale decision made before any fitting happened is usually the real reason it did.
Bookmark this valuable knowledge.
Peter Kempthorne, MIT professor whose volatility modeling course trains the quants who price every option on Wall Street:
"Model a stock price directly and every desk gets it wrong the same way. Model the log of the price instead, and the exact same math suddenly works. I ask every new student why, and most can't answer on day one."
this is the one modeling choice separating a pricing model that works from one that quietly fails.
strip away the notation and the reason is almost embarrassingly simple. a ten dollar stock and a five hundred dollar stock don't move by the same number of dollars on an average day, but they do move by roughly the same percentage.
model the raw price and you're implicitly assuming both stocks share identical absolute swings, which is false the moment you check real data.
play that forward to any risk system running today. this single substitution, log price instead of raw price, is the quiet foundation under nearly every option pricing model still in use.
most people staring at a price chart never ask which scale the model underneath it was actually built on.
nobody markets "we picked the correct scale for our variable" as an innovation. it sounds like a footnote, even though picking the wrong one silently breaks every assumption built on top of it.
The model gets the credit when it prices correctly. The scale decision made before any fitting happened is usually the real reason it did.
Bookmark this valuable knowledge.
Peter Kempthorne, MIT professor whose volatility modeling course trains the quants who price every option on Wall Street:
"Every basic pricing model assumes volatility is constant. I can show you the exact math proving that's false: today's shock predicts tomorrow's, and two parameters capture what a high-order model needs dozens to fit."
this is the two-parameter model quietly running under every modern risk system, and it replaced years of more complicated math.
strip away the Greek letters and it's a simple observation. Big price moves tend to cluster together, and calm periods tend to stay calm, at least for a while.
a model assuming constant volatility misses that completely. It treats every day as equally uncertain, when reality clearly doesn't work that way.
play that forward to any risk desk running live positions today. GARCH with just two parameters fits real volatility patterns about as well as an autoregressive model with ten or more.
most people watching a stock chart never notice the clustering pattern that's been formalized and taught for free since the 1980s.
the industry sells volatility forecasting like it requires massive models and proprietary data. the two-parameter version has been public and effective for decades.
Constant volatility was always the convenient assumption, never the accurate one. Two parameters were enough to prove it.
Bookmark this knowledge.
@Frandeeer bookmarked this, i agree with what Marco said, ngl
Choongbum Lee, MIT professor whose probability course trains the quants Wall Street actually hires:
"A casino only needs a 52 percent edge to guarantee the house wins, forever, given enough hands. Quant funds run the identical math on trades instead of blackjack, and the only real question is whether your edge survives contact with the variance."
this is the exact proof funds pay six figures for, and it's sitting free on a public MIT board.
strip away the trading floor mystique and it's just the law of large numbers. a tiny, consistent statistical edge becomes a near-certain profit once you repeat it across enough independent trades.
a single trade looks like noise. a hundred thousand trades with the same small edge stop looking random at all.
zoom out to how funds actually size positions, and this is the entire justification for trading small and often instead of big and rare. volume is what lets a real edge separate itself from luck.
most people scroll past this exact board and have no idea they just watched the actual math behind every quant desk on Wall Street.
this is exactly what gets buried under performance marketing. "We have a small statistical edge, repeated relentlessly" doesn't sound impressive on a pitch deck, even though it's the entire reason quant funds exist instead of gut traders.
A tiny edge is worthless on one trade. Repeated a hundred thousand times, it's the only thing standing between a fund and a casino player who just got lucky once.
Bookmark this
Jake Xia, former Morgan Stanley options trader who now teaches the math behind Wall Street at MIT and Harvard:
"Beating the index isn't a mystery, it's one linear regression. Run your returns against the market's, and what's left over after you strip out the correlation is the only part anyone actually gets paid for."
this is the exact regression sitting under every "our model beats the benchmark" claim in 2026.
the idea is simple. beta is just how much you move together with the market. alpha is whatever's left after you subtract that correlation out.
most of what looks like skill is actually just beta, correlated movement dressed up as talent.
here is why it matters right now. every AI benchmark comparison works the identical way: strip out what any baseline model would already get right, and what's left is the only number that means anything.
here is what gets buried. reporting the raw score is easy. reporting the score after removing what a trivial baseline gets for free is the number that actually separates a real edge from noise.
beating the index isn't hard to claim. proving the beating wasn't just beta is the actual test.
Bookmark and watch this.
https://t.co/ZxnvKqrScI
Andrew Ng, the Stanford professor who taught more of the machine learning field than anyone alive:
"Quant funds pay six figures for people who understand one tradeoff: exact and slow, or approximate and fast. I taught that exact tradeoff for free, and it's the same choice deciding whether a training run finishes this week or never finishes at all."
this is the tradeoff sitting under every training run and every execution strategy that actually ships at scale.
the idea is simple. one method scans your entire dataset before it dares take a single step toward the answer. It's exact, and it's slow.
the other method looks at one example, takes an imperfect step, then another, wandering the whole way instead of walking straight. It never lands exactly on the optimum. It just keeps getting close enough, fast.
here is why it matters right now. past a certain scale, hundreds of millions of rows, trillions of tokens, the exact method stops being an option at all. It's too slow to even take its first step.
the messy, wandering, never-quite-converging method wins purely because it moves. precision stops being the goal. speed becomes the entire strategy.
here is what launch posts and pitch decks skip. "Trained on trillions of tokens" is the headline.
the actual decision underneath, choosing an algorithm that admits upfront it will never be exactly right, in exchange for finishing at all, is the part that never makes the slide. it doesn't sound like a breakthrough. It sounds like giving up precision on purpose.
Nobody training or trading at real scale is chasing the exact answer. They're chasing close enough, fast, repeated a few trillion times.
Bookmark this.
Andrew Ng, the Stanford professor who taught more of the machine learning field than anyone alive:
"Quant funds pay six figures for people who understand one tradeoff: exact and slow, or approximate and fast. I taught that exact tradeoff for free, and it's the same choice deciding whether a training run finishes this week or never finishes at all."
this is the tradeoff sitting under every training run and every execution strategy that actually ships at scale.
the idea is simple. one method scans your entire dataset before it dares take a single step toward the answer. It's exact, and it's slow.
the other method looks at one example, takes an imperfect step, then another, wandering the whole way instead of walking straight. It never lands exactly on the optimum. It just keeps getting close enough, fast.
here is why it matters right now. past a certain scale, hundreds of millions of rows, trillions of tokens, the exact method stops being an option at all. It's too slow to even take its first step.
the messy, wandering, never-quite-converging method wins purely because it moves. precision stops being the goal. speed becomes the entire strategy.
here is what launch posts and pitch decks skip. "Trained on trillions of tokens" is the headline.
the actual decision underneath, choosing an algorithm that admits upfront it will never be exactly right, in exchange for finishing at all, is the part that never makes the slide. it doesn't sound like a breakthrough. It sounds like giving up precision on purpose.
Nobody training or trading at real scale is chasing the exact answer. They're chasing close enough, fast, repeated a few trillion times.
Bookmark this.
https://t.co/q9SBesYjPW
Ian Ball, MIT game theory professor:
"Citadel pays quants $400K to size a bet correctly under risk. I put the entire curve that governs that decision on a board for free: a guaranteed $10 beats a 1% shot at $1,000, and that's not caution, that's the math."
this is the free MIT board proof sitting under every position-sizing job that pays half a million dollars.
the idea is simple. nobody maximizes expected money, they maximize expected utility, and that curve bends. A safer bet with a lower average payout can be the objectively correct bet once you account for the shape of the curve, not the raw number.
here is why it matters right now. every fund that blew up chasing the highest expected value skipped this exact curve. the curve isn't a nice-to-have adjustment on top of the math, it is the math, and it's the same four-minute proof whether you're sizing a trade or scaling a training run.
here is what gets buried. expected value is one clean number, easy to put on a slide. the curve underneath it decides who's still solvent after the tail event actually happens, and it was formalized decades before any fund's risk desk existed.
Expected value gets the pitch. The curve decides who's still around to read the next one.
Bookmark and watch it today.
https://t.co/FgJC5WmYvy
Patrick Winston, MIT professor whose intro AI lecture trained the generation now running frontier labs:
"Every AI lab in 2026 is selling you the model as the breakthrough. I taught this course for forty years, and the model was never the hard part. The representation was."
this is the one slide every "our model changed everything" launch post in 2026 quietly skips.
the idea is simple. Winston refuses to define intelligence as an algorithm. He defines it as the representation underneath it the way you encode a problem so an answer becomes mechanical instead of mysterious. The algorithm is just what happens after that choice.
here is why it matters right now. every lab in this cycle ships a bigger model and calls it a moat, but the architectures are public, the weights leak, the papers get read by every competitor within a week. None of that was ever the scarce part.
here is what gets buried under the launch thread. The representation - the actual framing of the problem - is the part that doesn't demo well and doesn't fit in a keynote. Quant funds paying $500K for someone who can reframe a problem correctly know this. AI labs shipping another benchmark win mostly don't say it out loud.
The model gets the headline. The representation is still the whole subject, forty years later.
Bookmark this.
https://t.co/1APRwnKnAd
Last Seen Users on Sotwe
Trends for you
Most Popular Users
Elon Musk
@elonmusk
240.9M followers
Barack Obama
@barackobama
119.2M followers
Donald J. Trump
@realdonaldtrump
111.8M followers
Cristiano Ronaldo
@cristiano
111.4M followers
Narendra Modi
@narendramodi
107M followers
Rihanna
@rihanna
97.9M followers
NASA
@nasa
92.2M followers
Justin Bieber
@justinbieber
91.1M followers
KATY PERRY
@katyperry
88.1M followers
Taylor Swift
@taylorswift13
82M followers
Lady Gaga
@ladygaga
73.5M followers
Virat Kohli
@imvkohli
70.6M followers
Kim Kardashian
@kimkardashian
70M followers
YouTube
@youtube
68.7M followers
Bill Gates
@billgates
64.1M followers
Neymar Jr
@neymarjr
63.4M followers
The Ellen Show
@theellenshow
62.4M followers
CNN
@cnn
61.9M followers
Selena Gomez
@selenagomez
61.2M followers
X
@x
60.8M followers