There is a structural flaw running through the current generation of AI trading agents, and it is easy to miss.
Most AI trading agents can only execute one action: to buy.
But it's not for the reason you might think: almost every one of these systems is capable of opening a short position.
They choose not to.
And in a market that has spent most of the last decade trending upward, an agent that only buys looks like it is working right up until the moment it isn't.
What the research actually found
Several independent analyses have converged on the same result.
One study measuring signal distribution across LLM driven trading systems found a 90.4% long signal rate, attributing it to systematic bias in how these models reason about financial data.
Testing across models found the bias is not uniform but is often severe. One widely used model showed what researchers described as a strong and persistent bullish bias, consistently favouring longs and underweighting shorts, while another showed no consistent directional lean at all.
The most revealing experiment took a bullish-biased model and simply reversed every decision it made. The resulting short-heavy portfolio captured downside moves the original system had missed entirely. The signal was there but the model was structurally unable to act on it.
Why it happens
Three things compound.
Language models favour continuation. Given a price series and asked what happens next, a model trained on text tends toward trend-extrapolation. Mean-reversion and reversal require asserting that the recent past is about to stop being a good guide, making a weaker inference for a system optimised to predict the next plausible token.
The training corpus is directionally skewed. Financial writing about crypto is overwhelmingly bullish in aggregate. Analyst notes, project documentation, social posts, news coverage. The volume of text arguing that something will go up vastly exceeds the volume arguing it will go down. A model reasoning from that corpus inherits its slope.
Shorting is asymmetric and models are trained to be cautious.
A long position has bounded downside; a short has unbounded loss. Systems tuned to avoid confident, risky assertions will systematically under-produce exactly the signal that requires one.
None of these are bugs to be patched. They are properties of the approach, which is why the bias shows up across different models built by different teams.
Why a long-only agent isn't a strategy
Here is the part that matters for anyone allocating capital.
An agent that only takes long positions is not running a strategy. It is running leveraged beta, a bet that the asset goes up, executed with extra steps and a monthly fee.
In a rising market that is indistinguishable from skill.
Returns look strong. Backtests look strong. The dashboard is green. What is actually happening is that the market went up and the agent was long, which any buy-and-hold position would also have achieved, with lower fees and less operational risk.
The distinction only becomes visible in a drawdown. The exact moment you needed it to have been visible earlier.
An agent that cannot go short cannot protect you in a falling market unless it has an explicit flat state.
That last point is underrated. The ability to hold no position is as important as the ability to short. Many agents are structurally always-in-market, because a system built to find trades will find them whether or not conditions justify it.
Doing nothing is a decision, and most agents cannot make it.
What it takes to build an AI Agent that genuinely takes both sides
Three things, and they are architectural rather than a matter of prompting.
1. A decision cadence with a real flat state. The system must evaluate on a schedule and be permitted to conclude that no position is warranted. If the only outputs are long and short, the model will produce one of them regardless of conviction. Three outcomes can be choosen: long, short, or flat. This is what changes what the system is optimising for.
2. Risk sizing that is symmetric. If your short positions are sized smaller than your longs "because shorts are riskier," you have reintroduced the bias in the risk layer after removing it from the signal layer.
Downside protection has to be capped per trade in a way that applies identically in both directions.
3. Signal generation that is not purely generative. Systems relying entirely on a language model reasoning through market conditions inherit that model's directional slope.
Market-structure inputs: funding rates, open interest, positioning, do not have a bullish prior. They are just measurements and gives the agent the full picture rather than a biased take.
Win rate is the wrong question
Anyone evaluating a directional agent eventually asks how often the agent should take trades.
It is the natural question and it is the wrong one, in a specific and important way.
The point is fair: in a rising market, a high return proves very little, and a win rate near 50% proves less than it appears to.
But the conclusion people draw from it is wrong.
A directional strategy does not need a high hit rate. It needs asymmetry.
Consider a system that is right slightly more than half the time, caps its loss on every losing trade before entry, and lets winning trades run further than losing ones are permitted to fall. That system is profitable at a win rate barely above a coin flip but not because it predicts well, rather because the size of what it wins exceeds the size of what it loses.
Now consider a system that is right 70% of the time with no risk cap. One position moving violently against it erases months of accumulated small wins.
High hit rate, negative expectancy.
Win rate without the risk architecture tells you almost nothing. This is the same finding that shows up in retail trading data: most traders lose money, and the cause is rarely entry selection.
It is position sizing and the absence of a loss limit.
How Super Perps decides
Deploy's directional agent, Super Perps, is built around the three requirements above.
It reviews the market every four hours and decides long, short, or flat. The flat state is a genuine output, not a fallback. If conditions do not justify a position, it holds none.
Risk is capped on every trade before it opens. The maximum loss on any single position is defined at entry and applies identically to long and short exposure. This is what makes the asymmetry above work, and it is why the relevant number is not the hit rate.
It trades BTC, ETH, SOL and HYPE on Hyperliquid, liquid perpetual markets where positions can be entered and exited at size without moving the price against you.
The practical test
If you are evaluating any autonomous trading agent, the diagnostic is straightforward and most systems fail it:
- Ask to see the short trades. Not whether it can short, whether it has, and when. A decision log with no shorts in a market that has had down months tells you what you need to know.
- Ask what happens when it has no view. If the answer is that it always holds a position, it is always exposed.
- Ask how loss is capped per trade. If the answer is a portfolio-level stop or nothing at all, win rate is meaningless.
- Ask for something verifiable. A wallet address, an audited record, anything a third party can check. Screenshots are not evidence.
Most agents in this category answer the first question with silence.
Frequently asked questions
Why do AI trading agents rarely go short?
Three compounding reasons: language models favour trend-continuation over reversal; the financial text they are trained on is directionally bullish in aggregate; and shorting requires a confident assertion of unbounded-loss risk, which cautious systems under-produce. Research has measured long signal rates above 90% in some model-driven systems.
Is a long-only trading agent useful?
It is leveraged beta rather than a strategy. In a rising market it is indistinguishable from buy-and-hold with additional fees and operational risk. It cannot protect against, profit from, or step aside from a falling market.
Does a low win rate mean a strategy is bad?
No. Profitability comes from expectancy, not hit rate. A system right slightly more than half the time is profitable if its losses are capped and its winners are allowed to run further than its losers fall. A 70% win rate with no risk cap can still lose money on a single adverse move.
What does "flat" mean for a trading agent?
Holding no position. It matters because a system built to find trades will find them whether or not conditions justify it. An explicit flat state lets the agent decline to act, which is often the correct decision and one most agents cannot make.