TradeExpert, a trading framework that employs Mixture of Expert LLMs

I have been interested in algorithmic trading for quite a while now. Everything you've said resonates with me as I ran into similar issues. I hope you don't mind that I add my own couple cents here.

(a-c) LLMs are especially difficult to use due to their knowledge cutoffs and "unpredictability". A self-trained "old-school" machine learning model can go a long way though.

(d) With Crypto the volatility is great for trading, but liquidity can quickly become a problem (even at $1000 non-leveraged positions). For me, the ultimate goal is to find a strategy that is profitable in all market conditions. I personally value consistency and reliability more than absolute profit.

(e) There's some chatting about risk management, but absolutely no discussion on profitable strategies. Resources are incredibly scarce - Systematic Trading by Robert Caver is the only book that was actually useful.

I have zero knowledge of finance and trading, but when I got curious about algorithmic systems, it seemed like sentiment-based trading using current events was a more viable strategy than forecasting/regression-based analysis.

With cutting edge stuff you cannot say statements like "I tried X and found that it is not good at Y".

You could not get X to be good at Y, but it's not impossible someone else can.

> AI-based approaches are at best on par with traditional approaches

This seems like quite the generalization! Wouldn't it completely depend on the approach and model?

Are you referring to LLM's when you mention AI here?

Removing the "Market expert" which uses OHLCV (Open, High, Low, Close, Volume) also drops the sharpee from 5.01 to 1.88 while also increasing the max draw down to 13.29% (v.s. 9.70% for the index). I'd be very surprised if the pre training of the base model was the only source of leakage...

> 7B, Relatively small LLM.

Possibly too large a model. (Daily) finance data is finite - 7B parameters is potentially order(s) of magnitude more than the training data.

Its hard to find efficiency in markets when company values seem completely unhinged from reality.

Historically, companies were valued with a heavy weight put on their financials. That doesn't seem to be the case anymore, and without any clear approximation of how we are valuing companies it feels to me more like pure gambling.

I know that due to being interested in a specific industry, lets say Video Games, I can combine a general understanding of the environment, to a specific understanding of certain companies and the games they're making. If I see in a low-key mailing list that a company is making a new game that I know will line up well with the direction the industry is going, then I can invest in that company and likely be ahead of the market.

My hunch is that if you spend more than 6 hours a week studying the mood of a industry, you would likely be "luckier" than the market (although all risks still apply). I also believe a LLM could do exactly what I do with enough investment.

Markets aren't perfectly information-efficient. They're just _very_ information-efficient. So there's at least some room for non-luck-based strategies to outperform the market.

More important than information efficiency, is the subjective nature of investments, and the inter-sectoral complexity of the modern economy. There's lots of meaningful calculations you can do to predict the longterm value of a company, but massive and mostly unseen shifts in expectations happen regularly.

A fairly small number of people in Feb 2020 saw that 1/3 of students globally were suddenly home-schooled, and connected the dots to turn that into an ultra-strong signal to invest in Moderna.

Similarly, lots of people saw that Nvidia's CUDA granted it a near absolute-monopoly over anything that requires matrix multiplication. They were able to connect the dots to turn that into an ultra-strong signal that NVDA would turn from a relatively obscure company into a significant company.

Domain specialists have an edge over the aggregate-market within their narrow field of interest.

N=1, I think that:

- You'd have to be either extremely lucky, or spend a lot of money and effort (e.g. to get to the bleeding edge of quant or HFI game), to be in a position where you could use your knowledge and brainpower directly to give yourself an edge over the market, and:

- Any such edge is extremely short-lived - the moment you take advantage of it, the market itself will start adjusting to correct, and on top of that, other smart players will notice the irregularity, work to exploit it or reverse engineer your approach; in the end, the "pattern" quickly disappears.

Whether one can become able to continuously find such short-lived edges and profit off them, by means other than pure random chance, I don't know, but I highly doubt it. The space is way too competitive; a sustained miracle advantage would eventually attract regulatory attention, and/or unscrupulous parties willing to lie and cheat to bury you legally, or literally. But that's just me speculating.

IMO efficient market hypothesis is not true (assuming you have all available information you could beat it), but for an ordinary person there are so many ways the market is stacked against them that they should still just assume that it is true.

The question isn't "is it possible to outperform the market by thinking?", it is "will whatever advantage I can get by thinking weigh up for all the disadvantages I have from not being able to insider trade, being further from the exchange than other market participants, having lower ability to sustain losses than big market players, trading fees, taxes, opportunity cost, etc.?"

For any single individual, the answer to the second question is almost certainly no (unless you happen to sit in congress).

“In the short run, the market is a voting machine but in the long run, it is a weighing machine.”

Short term movements are very much sentiment driven, both humans and algorithms. Long term movements generally trend towards the consensus of fair value of any given security, which is usually modelled somehow. You can see stocks where the models struggle to reach consensus by their long term volatility.

I spent 6 years in algo trading. I think that the market is a thing people call the US stock market and it’s 98% machine traded volume. Whenever I hear people talking about their alpha on an equity I assume it’s either insider trading or ignorance of the billions in PhDs, super computing, and dark fiber they’re up against.

It’s pretty efficient at assimilating obvious revelations. You have to be quick to trade on the news, and you’re not going to find many arbitrage opportunities. There’s no reason why it would be particularly efficient at pricing long term outcomes. It reflects some sort of consensus or balance of opinions, but those things could simply be wrong. That said, even if you’re right about something, there’s so many factors to account for, it’s not going to be easy to profit. Personally I have most of my savings in retirement tracker funds (global equities), but I’m also building a portfolio of handpicked stocks/bets, aiming for 20, each with the same starting investment. Maybe it won’t be successful but at least it will be fun and minimise the regret of not acting on my insights/convictions. The diversification should stop it from being a disaster.

I stick with the dog walker analogy. A man walks down the street at a steady pace, and the dog on the leash runs back and forth and side ways and sits and jumps.

The man is the economy, the dog is the market. Unless you have access to HFT systems & models, there is no point in trying to guess where the dog will go. You have to look at the man, which is the aggregate health of the economy. If I start to see the man walk backwards, I sell, and vis versa. But *mostly*, I just stick things in ETF's and don't worry about it unless I need the money in the near future.

> Do you think the market is so efficient that anyone who outperforms it is merely lucky?

> Or do you think the market is inefficient enough for a person smart enough to be able to outperform it by thinking?

The market cannot be “efficient” (though this isn't what “efficient” means in economics besides in EMH), because if it was efficient, then there would be no incentives for market makers, and as such there wouldn't be anyone working to make it efficient.

As such, the market is always in an in-between state, inefficient enough so that people spend energy to try beat the average, but still hard to beat because there are already many players out there.

Sort of both. I believe there are massive polarized forces that invest heavily in understanding and gaming the system. Sort of like a big multi-player game of tug of war .

The system must be understandable or people wouldn't incorporate and collaborate on extracting an edge. Slack in the system represents a lower price point, which will be corrected by a corresponding purchase order.

An individual has an effect, but it's miniscule compared to the massive forces in play. Unless you have a TRULY novel analysis of a situation, you are going to have a very low probability, success rate and out competing the market.

I believe the markets are assymptotically efficient, but because of constant changes/surprises and the mechanisms being chaotic, it never reaches steady state, and thus is never actually efficient. There might be individual stocks or sectors that do reach steady state between annual reports, but as a whole: no. It's possible there's statistical signal in the inefficiency, comparing predictions to outcomes. At least HFT seems consistently profitable, and whatever Medallion is doing.

That said, judging if one specific trader is lucky or good is practically impossible.

If there’s volatility it’s not perfectly efficient, because corporate and economic fundamentals don’t change that quickly.

But playing volatility is a meta game. It’s about predicting what other players will do.

I have done, using my own real money, algorithmic swing and day trading for many years.

There are patterns that can be found and exploited, in the short term. However, I was unable to predict changes in the trading environment. My software did not even try to watch the news. So when big macro events happen, like a civil war in Syria, for example, the patterns can change overnight, so I sometimes was left holding a heavy bag.

Imo long term efficient, short term could be inefficient enough

I hear these all the time:

    1. Non-zero sum
    2. Market efficiency 
    3. Everything is priced in 

2 is true and actually a good thing for trading. It means there will be less noise vs signals.

3 doesn't matter if you are the market maker or you know the patterns caused by the "pricing in" and react to them.

The efficient market hypothesis also implies that it's impossible to consistently lose money for a given level of risk. Yet this happens all the time.

the markets are provably inefficient: companies shouldn't have their total marketcaps jostle by whole percentage points in a single trading day, when the underlying economics aren't that volatile.

but "markets" are (and have always been) "animal spirits" with all but participants leveraging their bets and other players pushing them off the table when they're slightly wrong.

With time and focus any person can beat the market. The market is wildly irrational, so rational actions are not the deciding factor. It is being able to spend all your time and focus watching the data. The question is whether this can be reliably outsourced to machines and what level of effort will it take.

A very simple exercise is to go and watch the live data for a single stock for 1 hour. You will see that it moves up and down and has trends. After 1 hour you can have an idea of an entry point, pretend to pick one, continue to watch, the odds are much greater than 50% that at some point the stock will go above your entry point in the next hour. It only doesn't if you bought into the top of a solid downward trend.

And that is the very basic example, but pair that with an understanding of options and a hedge strategy for the unlikely chance that you bought in at the top of a solid downward trend.

The last piece is to watch and understand Level II data and I reliably believe anyone can "beat" the market. The issue is that you literally have to monitor it consistently with a clear focus, can you do that with more than one security?

What portion can we reliably have machines do without some risk that the machine makes a tremendous error and blows up the account?

There is insufficient information in the market by design as evidenced by the existence of material nonpublic information.

Anyone who knows MNPI can make quite a lot of money which is why the SEC exists to regulate how that information can be used in the broader market.

To the extent that sufficient intelligence can reverse engineer or infer the private books and planning of public companies and the likely outcomes it can outperform other investors and make the market even more efficient.

Modeling global economies is hard; even regional models are hard and a large part why 20th century communism performed poorly compared to open markets (the others being internal corruption and external interference). Therefore I think individual humans have neither the capital reserves (bigger risks can be taken ~safely with more capital) nor the intelligence to beat the current markets. But it's not a zero-sum game and the global economy is still growing so if one's net outcome over several years is negative then it's probably due to bad strategy (e.g. picking stocks by hand, not diversifying) as opposed to bad luck.

If you have industry knowledge then you can outperform market short term

"Markets can remain irrational longer than you can remain solvent"

80% of it is the world's largest and longest-running Ponzi scheme. Because it's a Ponzi scheme, it brings negative expected value (maybe the other 20% cancels it out, maybe not). But because it's the world's largest and longest-running, you'd better participate because the chance it collapses at the exact moment you didn't invest but everyone else did is pretty low, and you can't let them outcompete you in the short term. You have a significant chance of not being the bottom tier of the Ponzi, unfortunately. If you do invest and it does crash, you're no worse off than everyone else.

That's for general market trends. Individual stocks also make no sense. You could buy and hold some very stable company's stock but you're making a relative pittance compared to just, like, changing jobs or something. Expected compound interest is not high enough to justify the risk of a catastrophic crash; also if you think a crash will happen any time in the next forever, you're better off holding cash or low-risk bonds until after the crash and jumping in then, as that will more than make up for the compound interest you make up on.

In the past, compound interest has been significantly higher than expected, and very much worth it, but only due to survivorship bias. There are many reasons to think the next 60 years won't be anything like the last 60.

I spend far too much time thinking about and building code for trading on betting exchanges, which isn't quite the same, but there are parallels.

Most experts agree that the hard definition of efficient markets (that all information is baked into a price immediately), doesn't hold, but the soft definition (that the price will veer towards true value), does hold. The big question is how quickly does it get there?

That means there is potential to make money (if it takes n minutes, and you are able to trade at n/2, and exit at n, you're going to make money). Insider traders can make the most money, but there is also money to be made on much longer trends too, I think.

Most of the super smart people throwing money at algos and hardware are specialising in HFT to try and trade at very high speed. They are typically not looking at much longer trends. That means there is perhaps more exploitable value for the individual in longer trend spotting compared to shorter term trends. (Note, I don't mean technical analysis when I talk about trends.)

Would I recommend you get into market making and HFT? Not without an eight figure sum to get started.

Would I recommend you get into value investing and looking at the long-term? Sure, just remember that diversification is good for a reason, and you might struggle to beat an index - most professionals do. If you enjoy it though, you might find it a good way to make your retirement savings grow, perhaps even a living.

Would I recommend you get into day trading? Probably not, but that doesn't mean there aren't successful day traders sat at home making a living trading on margin. It's hard work, and if you're smart enough to make it pay, you're probably smart enough to make more money doing something more productive for society instead.

I also don't think most people are smart enough to make money (important note [1]), if they try and work out by "thinking". Don't try and move from first principles. Learn from successful people, who aren't full of BS. The vast majority of people on YouTube and selling books are not successful. There are real success stories out there, you need to filter a lot to find them, but they exist. Use them.

[1] Your question was if people can "outperform" the market. You don't need to outperform the market, you don't need to be optimal, you "just" need to make enough money to meet your goals, and preferably more than you would get just from leaving your money in an index tracker. That's not the same thing. If you're trying to perfectly enter/exit trades and "beat the market", you're dead before you get started.

When Meta did a name change, another stock - Meta Platforms - saw a spike

I think if you're talking on portfolio level a lot of these things are used as signals and parts of a greater whole.

In that sense, they can indeed add value. My current project is a modern version of a classic Harry Browne portfolio with even asset allocations to gold/bonds/equity/commodities, with optional layers of sophistication according to spec needs.

Something like systematic macro could be analyzed as a standalone return stream, but it's more useful when considered as an input into allocation/leverage adjustments (ex: if geopolitical uncertainty readings are high, cut down the trend following exposure). Even the more robot/quant stuff like vol trading feeds back into the wider portfolio management and portfolio construction level to some degree.

If it did, I think it's safe to assume they would have sold it to a quant firm and we never would have heard about it.

It doesn't look like it. There isn't a clear comparison to a "human trader".

one drawback here is the backtesting. They backtesting only 1 year in 2023.

The phrase "changing the [something] game" is the most obnoxious.

We've been past that point for more than a decade. This is a gimmick.