Transcript
Use with ChatGPT / Claude Auto-scroll A
Explica por qué firmas como Susquehanna y Jane Street usan póker para entrenar traders, y presenta las cuatro ideas principales que se van a extraer de las charlas de MIT y Jared Akerman.
Poker is one of the most important cheat codes in trading, so much so that Susquehanna that manages over 800 billion dollars and had more than 20 billion in estimated revenue for 2025 from trading uses poker to train itstraders.And they're not alone.Jane Street includes poker in his trading internship and they're one of the biggest quant firms out there.Now, MIT actually built an entire course around this.Some of their guest lecturers include Bill Chen, a Susquehanna quant with two World Series of Poker bracelets, and Matt Lauerco, a former trader who won over million dollars in a single poker tournament.And there was Aaron Brown, a risk manager at AQR at the time.Now, you might be thinking, what does poker have to do with making money in trading?Well, a lot more than you'd actually think.I've taken four ideas from the MIT lectures and a Susquehanna quant's own explanation and turned them into a road map for you on how to think about trading using some of the poker principles.
Now, I'm going to include a poker example in this video that's going to change how you look at winning streaks in trading completely.Now, in the video from Susquehanna, Jared Akerman explains why they use poker to train traders.He worked in high-frequency trading for equities and options, and before that, he played poker professionally.He also co-wrote the book The Mathematics of Poker with Bill Chen, who's also in the MIT lectures.So, he's someone who can explain this from both sides, having actually done the work.Now, we're going to start with what the other person's actions tell you.Then we're going to look at how much evidence you actually have before you put money behind that read or trade.And then from there, we're going to go into building a strategy you can follow and then how to judge the results and handle the variance when you do everything right and still lose money.
Muestra cómo interpretar las acciones del otro jugador, estimar probabilidades y evitar conclusiones erróneas por sesgos o muestras pequeñas. También conecta esto con la selección de contrapartes y la selección de datos similares en trading.
So, Jared started by separating the information you have in poker because you've got your own cards and you've got the cards everyone else can see.And then you have the information you're getting from how the other players act.So, a big bet from someone who bluffs constantly definitely means something different from the same bet from someone who never bluffs.So, your chances of winning depends on who made that bet and how they play.Let's say we're on the river, so there's no more cards coming and there's $150 in the pot including their bet.You've got a hand that beats their bluffs, but loses whenever they're betting a strong hand.It costs you $50 to call, so you're risking $50 to win the $150 that's already there.Divide 50 by 200, and you get about 25%, which means you only need to win a quarter of the time to actually break even.
Now, let's assume player A is bluffing 40% of the time when they make this bet.You win $150 on those bluffs, and lose $50 on the other 60% of the time, which works out to about $30 of expected profit from this transaction.But, if player B only bluffs 10% of the time, that exact same call now loses $30 on average, even though your cards and the amount you're playing with hasn't changed at all.The difference is simply who you're playing against.Obviously, we're choosing those bluff rates, and for example, whenever in a real game you have to estimate them from what you've actually seen, which is where this gets a lot more interesting.
So, he comes along, and he wants to buy, say, some option, or say, some other product, and we, as market makers, are in the business of figuring out that that that thing that we're going to sell him isworth 305, when in fact, and then we want to sell it to him for 307, and make two pennies, and that's our business.And we can both do that trade.He gets to hedge his risk for not that much money, and we get to like make a small amount on a trade, um without a lot of risk.And so, like both sides can win to this.So,
potentially two cents of edge there before costs.The buyer might just be hedging a business risk, so paying a little extra for that protection can make sense for them.You can benefit without the other person necessarily being an idiot or making a bad trade.But, let's say the buyer knows something your model has missed, and accounting for the information puts the expected value at $3.10.You've sold it for 307, so what looked like an edge is actually an expected loss.And that's what adverse selection actually means here, because the trades you can get can be the ones you've priced incorrectly.Just like the bet in poker, someone's choosing to trade with your information, and your estimate has to account for that.To make this useful, you need some sort of idea of how the other players behave, but Jerry points out even a few hundred hands with someone covers very little of what actually happens in poker.
So, you also have to learn from other players who behave similarly, which gives you more information to work with.The difficulty is deciding what's similar enough because grouping completely different players together can give you the wrong answer.And if you want to help building a live trading system that's fully algorithmic, check out the Crypto Momentum group that's linked below.But, let's get back into how we group these players.If you put everyone in one group, you've got plenty of data, but you're averaging the aggressive bluffers together with the person who only bets with strong hands.So, that average might not describe either of them particularly well.But, if you build a separate model for every player, board, and bet size, you can end up with hardly any examples in each group.So, a couple of unusual hands completely change your estimate.What you're trying to do here is share information between situations that are similar enough to actually be useful.
And adding more detail to the model only helps if you've got enough evidence to actually support that detail.Jared makes the same point about recurring trades.What looks like new information might just be a trade that a lot of people tend to do at that time.For example, a scheduled rebalance and a reaction to news can both produce a buy order, but mean completely different things for the price.You can use observable information like timing to separate them, and then check whether those groups help your model on later data.
Usa una mano famosa de póker para ilustrar que una lectura fuerte no justifica una apuesta grande sin suficiente evidencia. Luego traslada esa idea a probar oportunidades de trading con tamaño pequeño y reglas aplicables en tiempo real.
Now, in MIT's decision-making lecture, Matt Lerchinko, the former Susquehanna trader and World Series of Poker winner, tells a story about someone who's desperate to reach the prize money.Matt thinks this player will fold almost anything, so he goes all in with four-deuce.Well, the guy says he knows Matt would do that and calls without looking, then turns over nine-deuce and wins.Matt was trying to predict his behavior while he was already anticipating what Matt would actually do.
So, I looked down and I see four-deuce offsuit.Um and we both have like medium stacks, but I have this hell of a read on this guy, right?So, I shove all in.He says, "I knew you were going to do that.I call blind."He calls without looking at his cards cuz he knew I was going to shove in.It's like the strongest read I've ever had.He flips over his hand.He has nine deuce offsuit.He has me dominated with a terrible hand.
Later in the lecture, Matt explains that a small amount of evidence calls for a small adjustment.Then you can consider moving further as evidence improves.Otherwise, you're putting too much behind a read that could be wrong.For example, if you spot a trading inefficiency, test it with a small size, check your feels and costs before adding to it, and you have to be careful around leverage here, because a profitable short run of trades canmiss huge losses or tail risks, and that edge can disappear very quickly as you scale up.You also need to use this while the game is happening, because you can't pause a hand to research the opponent.Jared's goal is to get your quick decision close to the one you'd make properly after studying it.MIT instructor Kevin Desmond explains this in his introductory lecture that you won't have all the information, a calculator, or much time to actually make a decision.
So, you need to practice rules you can apply.With an algorithm, you can research those rules beforehand and then let them apply them live on live trading data.
Presenta el juego de la moneda para explicar por qué una estrategia óptima necesita faroles y mezcla de acciones, no solo jugadas obvias. La sección enfatiza que una estrategia debe evaluarse por la respuesta que provoca en el rival o en el mercado.
Matt actually illustrates this with a coin flip game where both players put $1 in, and only you can see the coin.Heads means you've got the winner, while tails means you lose when it's revealed.
Each player antes a dollar.There are two players, and then you flip a coin.If it's heads, you win.If it's tails, your opponent wins.However, there's a round of betting.Only you see the coin after the flip.Then you can bet or check.So, you choose to bet a dollar or check, and your opponent can only check.If you check, they have to flip over their hand, or call.If you bet, they either have to call a dollar or fold.So, you know if you win or not.Your opponent doesn't know if they'll win at showdown.That's the idea of this game.
Now, the obvious thing here would be to bet whenever you get heads and check whenever you get tails, because then every bet you make is a winning hand.But as soon as the other player works that out, they can fold every time you bet, so you never actually get that extra dollar from that call.Your rule ultimately has made the decision incredibly easy.So, the balanced strategy still bets every head, but you also need some bluffs in there so the other player can't work out exactly when you're going to win.Let's say you bluff 1/3 of the time, which gives you one bluff for every free value of bets.So, that's about 25% of all bets are bluffs, making cooling and folding equally valuable for the opponent.But, let's say now they cool two out of the three times, which means your bluff wins $2 one out of three times and loses $1 two out of three times, averaging it everything out to zero.
With both players following those rules, neither can improve by changing only their strategy.Those bluffs affect how much you make with your good hands because they give the opponent a reason to cool.So, Matt's point here is to look at the whole strategy and the response it creates.For a market maker, changing the quote changes which orders trade with you.So, they have to look at the fills that rule produces.Subsequent prices and costs could be because they had to change their quote.Markets are all complicated, but you can't assume everyone else's response stays the same when you change your strategy.
Analiza cómo una pequeña ventaja puede quedar oculta por una enorme varianza, y por qué una sesión perdedora no prueba que una estrategia sea mala. Cierra con la importancia de registrar manos, revisar supuestos y mantener disciplina para no empeorar las decisiones después de perder.
Now, that number from the beginning comes from Jared's example of a heads-up limit hold'em, where a decent edge might be about 0.02 big bets per hand.A big bet is a larger fixed bet size in that game.
Okay, I'll give you an example from limit hold'em, a game that not played so much anymore, but in that game, in a heads-up game, you know, a pretty decent edge to have over an opponent would be like 0.02big bets per hand.Where the standard deviation of a particular hand would be 2.5 big bets per hand.
But, the standard deviation per hand is 2.5 big bets, which tells you how widely the actual results can vary around that tiny average profit.When you divide 2.5 by 0.02, that measure of noise is 125 times bigger than the edge.So, you can see why looking at whether one hand won or one hand lost tells you very little about how good that strategy actually is.Let's use these numbers and assume the hands are independent and the edge stays the same.After 100 hands, you'd expect to make about two big bets, but the standard deviation of the total is 25.So, there's a lot of uncertainty around that result.Using a simple bell curve approximation, that gives you a round of 40% chance of finishing down, even though that you have a positive edge.This is our illustration using Jared's input, but it shows you why a losing session on its own doesn't tell you whether your strategy is bad or good.
And if you're trying to estimate the average profit of a hand, the standard error with 100 hands is 0.25 big bets, which is still much bigger than 0.02's edge we're trying to measure.But, even with a 10,000 hands, that uncertainty only falls to 0.025.So, collecting 100 times more hands gives you 10 times the precision, and that's why small improvements are so hard to judge simply from profit alone, especially in trading where conditions can change so rapidly when you're collecting data.Now, there's also another thing to keep in mind here, because betting three times during the same poker hand doesn't give you three independent hands of evidence.You're still dealing with the same cards and the same outcome.It's the same when one trade gets split across 10 fills, because those aren't 10 independent tests of your signal.When Jared talks about making enough trading decisions to get closer to the long-run result, he's specifically talking about independent decisions.
And poker helps you get used to experiencing that variance, because our earlier call has $30 of expected profit, but still loses 60% of the time.Knowing the math doesn't necessarily make that really feel any better, but actually playing that out many, many, many times, you get used to that sort of variance.Jared describes poker as a training ground for exactly this, where you can make the right decision and still lose money.Going through it repeatedly gives you practice in handling that frustration without letting it decide how you play the next hand.
And so, poker is a nice training ground for you're going to do the right thing, and you're going to lose.And like, experiencing that, dealing with it, training yourself to say, "That's all right."And then, you can feel bad after the market closes, or you know, like, after the game is over, or whatever.But, on the next hand, on the next trade, you will be still making good decisions with your decision-making process, the way that you built it, and the way that it will make money.
Where it gets expensive though, is when you start calling weaker hands because you want the money back, or in trading increasing your leverage after a loss.That frustration hasn't created a better opportunity, but it has changed the decisions you're now making.Even with an algorithm, you could still interfere with the size or switch systems after a bad week.Having entry rules and risk limits set beforehand gives you something to follow while you can check whether losses fit your expectations, or whether something actually has changed.Obviously, you can't just call every loss bad luck because getting unlucky with the right odds is different from assuming someone's bluff 40% of the time when they were actually only bluffing 10% of the time.In that case, your estimate needs fixing.And that's the risk Jared points to when the other person knows something you don't because that trade might never have had the edge you expected.
Along side managing the price swings, you need to keep checking the assumptions behind the actual strategy.And that's why he put so much of an emphasis on talking about hands with other players because explaining what you thought the opponent had and why you acted can expose an assumption you hadn't questions.Jared also gets the MIT class to save their hand history so they can analyze those decisions.In trading, saving the model inputs and actual fills you do unless you work out what if anything should have changed.So, that's how it all connects because you are using the other person's action as information, learning from similar situations, and building a strategy that accounts for how people respond to it.Then you have to judge whether it's working and deal with the losing runs without making the next decision worse.Poker gives you practice with all of that against people trying to beat you, which is why it makes it such a good training ground at all these big firms.
And if you want help building your first live algorithm, the code, the templates, and the guidance, check out the Crota Momentum Group, which is linked below.Let me know which of these four you need to work on most, and I'll leave the MIT course and Jared's talk in the description below.I'll also leave a video around how to backtest so you can get further into testing your strategy and understanding if it actually works live.