A Negative-Sum Game, Being Reshaped by AI

Short-term trading is probably one of the most alluring and yet least understood corners of finance. First, it is really hard to resist trading after seeing a bull market spanning from one year to another, with retail platforms marketing trading as easy as tapping on a phone; brokerage advertising features that leverage up investment; influencers post life-changing wins on meme channels. However, the mathematics tells a less flattering story.

Trading is negative-sum by nature

In 1991, William Sharpe published a one-page proof that has never been refuted. Divide every investor in a market into two groups: passive investors of the market portfolio (buy and hold) and everyone else (active investors); an accounting identity follows. Passive investors earn the market return by definition. Active investors, taken together, must also earn the market return, because together with passive investors they are the market. Before costs, the average active dollar earns exactly what the average passive dollar earns. There is no escape from this; it does not depend on efficiency, rationality, or skill. Once costs enter the picture — fees, spreads, commissions — the average active investor must underperform passive by precisely the cost difference. Trading is not zero-sum. It is negative-sum, and the more frequently one trades, the more negative it becomes.

Most lose, only a few win

What this aggregate identity hides is a striking distribution underneath. Empirical work on retail short-term trading — most rigorously by Barber and colleagues using complete Taiwan exchange records, and by the Brazilian regulator using Brazilian retail data — shows that returns are not normally distributed around the negative mean. They are heavily right-skewed with a long thin tail of winners and a heavy mass of losers clustered slightly worse than the average. Of day traders who persist longer than a year, roughly 97 percent lose money net of costs. Less than one percent earn enough to call it a living. The top 500 traders in the Taiwan study captured roughly the same total profit as the next 4,000 combined; the bottom 400,000 funded the difference.

The participant base sorts naturally into cohorts: a sliver of market makers and HFT firms collecting consistent profits from the spread; a few percent of institutional prop traders running with a slight edge; perhaps one to two percent of retail traders with persistent skill; a recreational middle that loses gently and steadily; and a long tail of active day traders who lose significantly. The median experience sits left of the mean, which is itself left of zero. Short-term trading is, in shape, a negative-mean lottery — the worst combination, because even the right tail is too thin to compensate. This makes short-term trading a striking resemblance to playing poker in a casino: the house always wins no matter what.

AI may not be of help

Into this negative-sum game, AI agents are now arriving in volume. They will not break Sharpe's arithmetic. The aggregate must remain negative-sum after costs. What AI changes is the shape. The right tail will compress and concentrate: the few winners will become fewer and bigger, as edge migrates toward firms that can afford the best foundation models, the most data, and the deepest engineering benches — the same pattern HFT enacted on floor trading between 2005 and 2015. The middle of the distribution will homogenise. Mid-tier firms and ambitious retail traders deploying AI tools will draw on the same training corpus and the same handful of foundation models, producing strategies that correlate more than independently designed quant systems ever did. This is the crowding impact we explored in another article. The mean for this group will barely shift, but their losses will synchronise — clustering in time during stress events and fattening the extreme of the left tail.

The recreational middle will erode as AI tools accelerate retail traders' descent through leverage and complexity rather than lifting them toward profit. The "rake" itself will grow: cloud compute, data vendors, model APIs, and AI infrastructure providers add new mouths to the intermediation chain, dragging the mean further left. And reflexively, as AI agents read the same crowding research and learn to avoid the obvious traps, hidden concentration will migrate to less monitored corners — exotic options, cross-asset hedges, and other corners regulators do not yet track. The system will look smoother in normal times and break harder during stress.

A short summary

None of this is hypothetical mathematics. It is a forward extrapolation of mechanisms already documented in the crowding and microstructure literature. Sharpe's identity guarantees the aggregate outcome; the distribution determines who bears it. AI does not spread winning; it concentrates wealth generation. The few winners will be richer and fewer. The many losers will be more numerous, more correlated, and more catastrophic when the next stress event arrives. The recreational middle ground will likely disappear. The infrastructure vendors and the firms with the largest models will capture an expanding share of an unchanged negative sum. Anyone considering short-term trading in the era of AI agents should ask not whether the game is winnable in principle — Sharpe answered that — but whether they have any reason to believe they sit in the part of the distribution that AI is making smaller.

Note on terminology

"Negative-sum" in this article refers to the aggregate of active trading relative to the market return, not to absolute money outcomes. Sharpe's identity says the average active investor underperforms the average passive investor by precisely the cost difference. In a rising market, both groups can still make money in absolute terms; the active group simply makes less. In a falling market, both lose, and the active group loses more. The "loss" is the gap to the passive benchmark, not necessarily a loss of capital. The 97 percent retail loss statistic discussed in the article reflects something more severe — short-term retail traders combine the Sharpe drag with adverse selection from informed counterparties, leverage, and poor timing, which together turn a relative underperformance into absolute capital destruction.