AI's Effect on Markets – Composer

AI's Effect on Markets

AI will make up over 10% of trading volume in two years. Yet the best human traders will earn more than ever. What gives?

I have two seemingly contradictory predictions about how AI will reshape capital markets two years from now:

The agents are coming.

Machines already rule trading, even if most retail traders are blissfully unaware: the IMF estimates that ~70% of US equity trading is executed by algorithms rather than human traders. Having AI - or autonomous AI “agents” - trade on behalf of humans is a natural evolution of algorithmic trading.

Instead of having humans parse through mounds of data and hand-code algorithms, we will see humans delegate much of the heavy lifting to AI agents that specialize in trading. The trading agent will query various data sources, construct a trading strategy (expressed as an algorithm under the hood), and update its strategy in response to feedback signals from the market, all in an autonomous, end-to-end loop.

We already see early signs of retail’s appetite for trading agents, with viral Reddit posts and threads on X about people leveraging ChatGPT or Perplexity to gain an advantage on the market. These attempts are mostly hacks - relying on DIY solutions that look like toys.

Considering that Composer was born from my own DIY trading system, I’ve learned to never dismiss what might look like a hack or a toy. As Chris Dixon, general partner at a16z, famously put it: the next big thing will start out looking like a toy. Companies like Composer will be putting in the work to smooth over the rough edges of trading agents, turning them into polished, vertically integrated products.

As trading agent products get better, the market should grow, attracting even more attention and investment in the space, all in a self-reinforcing cycle. This is why I conservatively forecast that at least 10% of trading volume will be executed by trading agents in 24 months. The exact percentage is hard to forecast, but this was the lower end of our back-of-napkin estimate.

AI won’t solve the market.

Some believe they can prompt their way to becoming the next Jim Simons: “Hey ChatGPT, generate alpha for me.” This is structurally impossible – and anyone who trades for profit should be grateful.

First, any strategy – AI or otherwise – will face capacity constraints: even Jim Simon’s Medallion Fund, arguably the best performing fund in history, has been closed to outside investors since 2005 because excess capital erodes returns. A public AI agent can’t restrict the volume of assets following its trades; any edge would evaporate as adoption grew.

Even putting aside capacity limits, AI would not have the kinds of original insights that would generate alpha in the first place – at least if it’s absent human input.

Why? It’s because the foundation models the trading agent will inevitably rely on are explicitly trained to produce central answers.

We see it everywhere today: the sterile blah-ness of inoffensive content clearly generated by AI, devoid of any perspective or soul. Yet as Howard Marks has long argued, you outperform only by holding correct, non‑consensus views. The path to alpha is inherently at odds with AI models tuned to median preferences.

Some degree of market inefficiency is desirable, necessary even, for markets to function: as Joseph Stiglitz put it decades ago, markets have an “equilibrium degree of disequilibrium.” If all information is perfectly reflected in prices, then traders with an edge would stop trading, in turn leading to markets being informationally inefficient again, attracting traders once again, and so on. Likewise, If AI made markets perfectly efficient, traders would have no incentive to trade.

Fortunately for humans, perfect efficiency seems as untenable as ever. While AI will do a remarkable job at compressing and organizing information, it will also be a massive generator of noisy information to process in the first place.

One of my favorite memes is the notion that someone sends some bullet points to AI, the AI turns the points into an article, then the receiver summarizes the article back into a set of bullet points. This is an apt metaphor for markets as well: AI will lead to more company formation, more information creation, and more macro instability - all of which we will need AI to contend with. AI is the ultimate entropy wrangler – but it is also the ultimate creator of entropy. The net effect on market efficiency is likely to be totally offsetting.

AI is an amplifier of human ingenuity

So why am I so excited about trading agents, if the AI models themselves remain idiot savants at their best and midwits at their worst? Because an infinite arsenal of ansavants at a clever person’s disposal creates an almost unimaginable degree of leverage, to put it mildly.

With support from AI, you will get to spend all of your intellectual energy on the “what” – what you ask for and what you reject. The agents you command are like hyper-specialized employees who are masterful at figuring out the how– better, in fact, than even the most exacting humans.

In the next 24 months, agents will be able to handle increasingly long-running tasks, and their ability to use tools will accelerate through advances in reinforcement learning. In the context of trading, this means an agent will be able to sift through all publicly available information, including anything on the internet, time series data, and any alternative data sources it has available. The agent will then be able to compress and extract the signal from this data, passing it off to a superhuman-level coder to turn into a trading algorithm.

It will be the human, however, that articulates a set of priors, steering the model to pay attention to some signals and not to others, and to shrink down the strategy search space into something far more opinionated. The model will present ideas, of course, but the human will apply their taste to accept or reject these proposals.

We already see how human talent must collaborate with AI in creative domains, such as image and video generation. I have yet to see a compelling AI video created by an untalented person. The best AI videos are still labors of love–the result of many takes, with most generations discarded. Trading agents will similarly reward the most creative and passionately obsessive.

AI’s effect on the distribution of trading profits will be split: the best traders will earn more than ever, while median traders may very well go extinct. We see this dynamic already playing out with software engineers: the best receive $100M offers from Meta, while new grads struggle to find jobs at all. While AI magnifies inequality, it also knocks down barriers: with the right tooling, a lone retail trader could deploy the same tactics as large hedge funds.

Trading agents will require an entirely new interaction paradigm for human-agent collaboration. How do we ensure that the agents' decisions are totally transparent to humans? How do humans interact with the agent once it's up and running autonomously - how and when do we intervene? And how do we make sure that humans get to maximally impart their strengths – their taste, judgement, and conviction – at just the right times?

User experience design in a world filled with agents is going to be the product design challenge of the coming years in all domains, not just trading. Composer’s philosophy since founding is to design products that let customers focus on the art of trading rather than the drudgery. Our approach to trading agents will continue in the same spirit–designed to augment, rather than replace, human skill.