In the last post we walked through the Alpha Arena results: six frontier LLMs, four of them losing 60-plus percent. The obvious reflex looking at data like this is to ask "why doesn't the AI work yet?" — as if this were an implementation problem to be fixed by the next model generation.
That's the wrong question. The right question is one layer deeper and more uncomfortable: Why should language models (or any models) be able to see the future when no one else can?
This post is an attempt to unpack a piece of intellectual common sense that's getting trampled in the AI trading hype: markets aren't a problem you solve with more compute. They're something far more interesting.
First truth: nobody sees the future
Before we talk about LLMs, about hedge funds, about discretionary traders or algos: has anyone in the last 200 years proven they can reliably predict markets?
The answer is: not in any strict sense, no.
There are traders and funds who've achieved consistent positive returns over decades — Renaissance Technologies' Medallion Fund, Jim Simons, a few others. But even these exceptions don't sell their edge. Medallion has been closed to outside investors since the 1990s. If you had a reliable method to predict markets, you wouldn't license it for $99 a month — you'd use it yourself and not write about it.
That's not cynicism, it's economic logic. Profitable market knowledge gets destroyed by distribution. When five people know the strategy, it keeps performing. When five hundred thousand know it, it's gone, because everyone does the same thing simultaneously and there's no one left to take the other side. The only edges that survive are the ones that don't get widely distributed.
A strange property of AI trading product marketing follows from this: the marketing itself is proof the product doesn't work. If QuantPilot, BotPredictAI, or any other AI tool could actually see the future, the providers would have two rational options — get rich themselves, or sell it as a single client to one hedge fund for $100 million. Selling it as a $99/month subscription to retail would be the economically most senseless option. The fact that it's happening is the fact the product doesn't deliver what it promises.
Second truth: why do markets exist at all?
Here it gets philosophically interesting. Imagine all market participants were rational. All had the same information. All processed that information perfectly.
Then the price of every asset would have to be correct at every moment — it would reflect exactly what we collectively know about the asset. But then no one would have a reason to trade anymore. Whoever buys would think it fair at the same price the seller does. There'd be no movement, no incentive, no trade.
Markets would cease to exist.
That's not my observation. That's the Grossman-Stiglitz paradox, formulated in 1980 in one of the most influential papers in modern financial economics: "On the Impossibility of Informationally Efficient Markets." Sanford Grossman and Joseph Stiglitz (the latter would later win the Nobel Prize in Economics) showed mathematically that perfectly efficient markets are logically impossible. If prices reflected all information, no one would have an incentive to gather information — because gathering information costs something, and the reward would be zero since prices are already right. Without information gatherers, prices would become inaccurate again. It's a logical knot that doesn't untie.
Grossman and Stiglitz's solution: markets exist in an "equilibrium degree of disequilibrium." Prices reflect information partially but never completely, so that informed participants have an incentive to keep gathering information, which then gets partially incorporated into prices.
This has a beautifully paradoxical implication: Markets only work because they don't work perfectly. They're not truth machines. They're mechanisms for the ongoing, never-ending negotiation of disagreement between participants with different information, different interpretations, and different time horizons.
Third truth: every trade has a loser
From the Grossman-Stiglitz paradox follows something systematically suppressed in crypto marketing: at every single trade, one side is convinced it's smarter than the other. And on average, both are wrong.
Concretely: if you buy BTC, someone else sold BTC. That person isn't crazy. They thought the current price was high enough to sell. You thought it was low enough to buy. One of you will turn out to be right in hindsight — the other won't. Aggregate market statistics suggest both sides are right about half the time, which means the expected value of your disagreement before costs is roughly zero. After costs — spread, slippage, fees — it's negative.
Whoever "consistently makes money in markets" makes money against someone else who "consistently loses money in markets." It cannot be otherwise. In crypto spot, the typical distribution is: retail loses systematically, market makers earn the spread, large institutionals earn through edge strategies (information, speed, capital), MEV bots earn through technical advantages. When you trade as retail, the math isn't on your side. That's not a conspiracy, it's statistics.
Now consider what happens if an AI tool actually had a small edge and was sold en masse. A hundred thousand retail traders use the same tool that tells them "buy BTC now." They all buy simultaneously. Who sells? No one from the group of tool users — they're all buying. So market makers and institutionals sell to them at a slightly higher price. That's exactly the opposite of the promised edge. The edge gets converted by distribution into an anti-edge.
AI trading tools mathematically cannot produce mass-market edge. If the tool actually works and is sold to many, it loses its edge through distribution itself. If the tool doesn't work, it never had an edge. The only scenario in which a widely distributed AI tool could durably deliver edge to its buyers would be that the buyers never use it enough to influence market dynamics — meaning most buyers don't actually use their subscription.
Observably, that's exactly what happens in every newsletter, trading course, and signal service industry. They sell promises whose fulfillment they can't afford.
Fourth truth: what backtesting actually does
Here we get to why this post sits on a backtesting platform and not on a philosophy blog. If no one can see the future, what's the use of backtesting?
The honest answer is: backtesting doesn't claim to show the future. It shows what worked in the past — and that's a much more modest, but much more honest claim than "this strategy will work."
What backtesting actually delivers:
It eliminates strategies that wouldn't have worked in the past. If a rule consistently loses across 10 years of BTC history, it's not a promising candidate for the future. That's no guarantee for the future — but it's a strong indication that the assumptions behind the strategy aren't stable.
It calibrates expectations. If a strategy historically delivered 12% CAGR at 30% max drawdown, those are the realistic expectations for the future — with the caveat that the future is never like the past and max drawdown in live trading is typically larger than in backtest.
It forces discipline. A clearly stated, tested rule resists discretionary impulse. Someone with "buy BTC at RSI < 30 on 1W, respect SMA-200 trend" as a rule makes better decisions in drawdown than someone without a rule.
It makes assumptions explicit. A tested rule forces you to write down your assumptions about markets. "I believe pullbacks in strong trends offer profitable entries." You can then hold that against data. AI bots don't force you to assumptions like that — they just claim to have the right answer without you knowing it.
What backtesting doesn't deliver — and no honest backtesting platform claims it delivers:
It doesn't predict the future. Nobody can. Backtesting isn't an oracle.
It doesn't guarantee outperformance. When a strategy worked historically, that could be chance, could be a real edge, could be an edge that existed in the tested regime and won't exist in future regimes. Out-of-sample tests and multi-regime tests reduce that risk but don't eliminate it.
It doesn't protect you from yourself. A tested strategy only works if you implement it exactly live. Most traders deviate from their tested rules under stress, which makes all backtest statistics meaningless.
Fifth truth: why this matters right now
The AI industry in 2026 is selling a promise that resembles its 2024 promises but is one level stronger: not just "AI understands language" or "AI writes code," but "AI acts autonomously in the world for you." Trading bots are the most prominent manifestation because the promises can be measured directly in dollars.
The promises are being empirically refuted — Alpha Arena is the clearest data point so far — but that doesn't stop the industry from making them. On the contrary. The clearer it becomes empirically that autonomous AI agents don't work, the more expensive the lifetime packages and premium tiers selling the promises become.
That's a predictable economic dynamic. The consumers of trading tools are mostly not in a position to rigorously check the promises — they have no quant finance background, didn't read Grossman-Stiglitz in school, no experience with out-of-sample tests. They buy narratives. And narratives get more expensive the better they are, not the truer they are.
The intellectually honest position is less romantic: nobody can see the future. Markets aren't a solution to this problem, they're its institutionalized expression. Backtesting isn't a tool that reveals the future, it's one that holds strategies against the past and makes explicit what you actually believe. This position can't be sold as a lifetime subscription because it contains no promise you'd want to buy.
It's still the right position.
Sixth, final truth: what you can do with this
If you've read this far, the likely question is: "Okay, nobody can see the future, markets are paradoxical disequilibria — what do I do then?"
Three recommendations following from the above:
First, treat trading like insurance, not like prediction. Instead of asking "which strategy will win?", ask "which strategy has acceptable losses in the scenarios where it loses?" That's a different mental frame. You calibrate risk, you don't predict the future.
Second, accept that your edge — if you have one — comes from something others don't have or don't use. That can be information (rare, mostly illegal), patience (more common but underrated), discipline (very rare), or access (e.g. to institutional markets). If your "edge" is what a $99 tool promises you — everyone has it, so it isn't one.
Third, use backtesting for what it's for: making your own assumptions explicit and holding them against historical data. Not to predict the future. To see whether your assumptions about the past even hold up. If yes, you have at least a consistent basis for your expectations. If no, you know it now and before you've lost real money.
Backtesting Arena is built with this stance. It's a tool that makes your hypotheses testable and gives you realistic drawdown and performance expectations. It's explicitly not a tool that tells you what to do. That separation isn't a feature gap, it's an intellectual stance we won't change.
If that's too unromantic for you and you prefer to listen to someone who tells you the next AI generation will crack markets — the supply for that is currently especially plentiful. But before paying money for it, read the Grossman-Stiglitz paper. Or at least the first two pages. It costs nothing and hasn't gone away since 1980.
FAQ
If markets aren't efficient, then edges do exist? Yes, but transient and mostly not scalable. Grossman-Stiglitz says there must be edges, otherwise there'd be no markets. But they're distributed: information gatherers get a small edge for gathering information, the edge gets partially built into prices, others must gather further information to find the next edge. That's not the same as "buy this tool and the edge is yours."
Does that mean trading is always gambling? No. It means trading is a skill game with a high chance component. Like poker. In aggregate, skilled poker players systematically beat unskilled players. But every individual hand is chance-dominated, and short-term outcomes are poor indicators of skill. Trading has the same structure. That's what makes backtesting valuable: it gives you hints about whether your "strategy" is even one, or whether you just had a lucky streak.
What do you think of Renaissance Technologies / Medallion Fund / Jim Simons? Honest answer: there are very few demonstrably consistently profitable systematic funds in history. Medallion is one. They sell nothing to outside investors. They don't publish their methods. They're one of the intellectually most honest quant firms — and they're the exception that proves the rule: if you have a real edge, you keep it to yourself.
You criticize AI trading tools — why is your platform allowed? We criticize AI tools that make trading decisions or predict. Backtesting Arena does neither. We offer tools for strategy validation, which users feed with their own strategies. The decision is the user's. The strategy comes from the user. We validate it honestly against historical data. That's a fundamentally different category from "AI trades for you."