Reversal is an experimental transformation, not investment advice. Review the simulated competition and risk rules on How It Works.
Does Doing the Opposite of AI Trades Work?
It worked for three of four reverse configurations in one 28-day season. Reverse DeepSeek, Reverse Claude and Reverse Qwen swept the podium; Reverse Kimi lost 9.27%. The result is evidence that an inverted signal can outperform under some conditions, not that AI recommendations are generally wrong or that reversal will transfer to another market.
Why We Built Reverse AI Trading Bots
What if the models are *consistently* wrong?
Not randomly wrong. Not occasionally-bad-call wrong. Consistently, directionally, stubbornly wrong in a way you could actually exploit. That question led us to build what might be the first systematic reverse AI trading experiment, a contrarian AI trading strategy that bets against AI predictions.
We had been running TradeRank.ai, a live AI trading arena, for several weeks. AI models from OpenAI, Google, Anthropic, xAI, DeepSeek, and others trade crypto with simulated capital. We had performance data on every model. We knew their biases. One pattern kept showing up: some models were perma-bulls. Every cycle, every market condition, they wanted to go long.
So we asked a simple question: what if we took those models, listened to every trading decision they made, and did the exact opposite?
We built reverse adapters for four different AI models, deployed them into the same competition with real market data, and let them trade for two weeks alongside the originals. The results were more interesting than we expected.
This article is for educational and entertainment purposes only. It is not financial advice. The trading results described are from a simulated competition with no real money at risk. Past simulated performance does not predict future results.
Data as of: Day 14, February 22, 2026. Updated results available in Contrarian Agents Swept Season 2 (final results).
The core idea: if an AI model has a consistent directional bias (always bullish or always bearish), you can build a profitable strategy by systematically inverting its output. We tested this with four models. Two made money, one showed a marginal improvement, and one lost even more.
How the Reversal Actually Works
The mechanism is straightforward. Each reverse adapter wraps a real AI model (DeepSeek R1, Claude Haiku 4.5, Qwen3, or Kimi K2). It sends the same market data and prompt to the base model. It gets back the model's genuine trading decisions. Then it flips every directional trade.
Every `open_long` becomes an `open_short`. Every `open_short` becomes an `open_long`. Closes and holds pass through unchanged. Closing a reversed position is still the right thing to do, and holding means holding.
The tricky part is stop losses. When a model says "go long at $100 with a stop loss at $97" (3% below), the reversal needs to place a short at $100 with a stop loss at $103 (3% above). The adapter estimates the entry price from the stop loss distance and mirrors it to the other side.
The Core Inversion Logic
protected invertDecision(decision: CompetitionDecision): CompetitionDecision {
const inverted = { ...decision };
switch (decision.action) {
case 'open_long':
inverted.action = 'open_short'; // Long becomes Short
inverted.reasoning = `[REVERSED] ${decision.reasoning}`;
if (decision.stop_loss_price !== undefined) {
// Mirror stop loss to the other side of estimated entry
const estimatedPrice = decision.stop_loss_price / 0.97;
const slDistance = estimatedPrice - decision.stop_loss_price;
inverted.stop_loss_price =
Math.round((estimatedPrice + slDistance) * 100) / 100;
}
break;
case 'open_short':
inverted.action = 'open_long'; // Short becomes Long
inverted.reasoning = `[REVERSED] ${decision.reasoning}`;
if (decision.stop_loss_price !== undefined) {
const estimatedPrice = decision.stop_loss_price / 1.03;
const slDistance = decision.stop_loss_price - estimatedPrice;
inverted.stop_loss_price =
Math.round((estimatedPrice - slDistance) * 100) / 100;
}
break;
case 'close':
case 'hold':
inverted.reasoning = `[REVERSED] ${decision.reasoning}`;
break;
}
return inverted;
}The market analysis gets annotated too. Every reversed response is tagged with `[REVERSED]` in the overall assessment and `[CONTRARIAN]` in the risk notes. When you read the reasoning on the live site, you can see exactly what the base model wanted to do and what the reversal actually executed.
One more detail that matters: the reverse adapter also handles the two-tier scanning flow. In Season 2, with 89 tradeable assets, models first scan the universe to pick 3-5 interesting assets, then make detailed decisions on those. The scan phase passes through without inversion. We only flip the actual trade decisions, not the asset selection. The model picks what to look at; we just disagree about which direction to trade it.
The Results: A Story That Splits in Two
We deployed reverse adapters for four models in Season 2 of the TradeRank.ai competition. Each reverse agent started with $10,000 in simulated capital and traded the same 89-asset universe (equities and crypto) as everyone else. The base versions of these models had been frozen mid-Season 2 when they were removed from active competition. We knew their returns and biases up to that point.
Here is what happened.
Reverse Agent Performance vs. Base Models
| Base Model | Base Return | Reverse Version | Reverse Return | Spread |
|---|---|---|---|---|
| DeepSeek R1 | -3.39% (frozen in S2) | Reverse DeepSeek | +2.64% | +6.03 pts |
| Qwen3 | -1.63% (frozen in S2) | Reverse Qwen3 | +1.18% | +2.81 pts |
| Claude Haiku 4.5 | -3.26% (frozen in S2) | Reverse Claude | -2.70% | +0.56 pts |
| Kimi K2 | -0.76% (frozen in S2) | Reverse Kimi | -6.70% | -5.94 pts |
Two reversals clearly worked. One was marginal. One made things dramatically worse. The spread between best and worst outcomes is nearly 12 percentage points: +6.03 for Reverse DeepSeek to -5.94 for Reverse Kimi.
This is not a story about contrarian strategies being universally good or bad. It is about *which models* you choose to bet against, and why. For the full leaderboard and analysis, see We Made 13 AI Models Trade Against Each Other.
Deep Dive: Why Reversing DeepSeek Works
Reverse DeepSeek is the star of this experiment. It returned +2.64%, finishing near the top in a competition of 13 models. As of this Day 14 snapshot (February 22, 2026), it had just reached its all-time equity high. View Reverse DeepSeek's competition performance in the Season 2 archive.
The numbers tell a clear story.
Reverse DeepSeek Stats
| Metric | Value |
|---|---|
| Return | +2.64% |
| Total Trades | 65 |
| Win Rate | 35% |
| Short Preference | 76% of positions |
| Best Trade | +$209.15 shorting OPUSDT (20.15% gain) |
| Max Drawdown | 2.24% |
| Total Fees | $73.64 |
Why it works: DeepSeek R1 is a reasoning model. It "thinks" through its decisions with extended chain-of-thought. In the current market environment, that deep reasoning consistently concluded that assets were undervalued and due for a bounce. The base DeepSeek deployed 85%+ of its capital into long positions early in the competition and kept piling in.
When the market turned choppy and sideways, those longs bled. DeepSeek R1 was frozen at -3.39% when removed mid-Season 2.
The reverse adapter took every one of those bullish calls and turned them into shorts. Since the base model was 85%+ long, the reverse version ended up 76% short. In a choppy, mean-reverting market, shorts on overextended rallies worked. For detailed model-by-model comparison, see GPT-5 vs Claude vs Gemini vs Grok.
The best single trade illustrates the inversion. Reverse DeepSeek shorted OPUSDT (Optimism) at $0.1618 and covered at $0.1292 for a $209.15 profit, a 20.15% gain on a single position.
“[REVERSED] Locking in 20.15% profit as 4h RSI at 11.9 indicates extreme oversold conditions and high probability of bounce.”
Read that reasoning carefully. The base DeepSeek saw RSI at 11.9 (deeply oversold) and wanted to buy the dip. Classic textbook reasoning. The reverse adapter shorted instead. And the price kept falling.
The core insight: DeepSeek R1's reasoning is *internally consistent and well-argued*. It just happens to be consistently wrong about direction in this market regime. The model builds a compelling case for why an asset should go up, marshaling technical indicators and pattern analysis. The market does not care about compelling cases. When the prevailing trend is choppy and downward, being a well-reasoned bull means being a well-reasoned loser.
A 35% win rate might look bad, but the winners were significantly larger than the losers. Reverse DeepSeek let its winning shorts run while the mirrored stop losses cut losers quickly. The asymmetry of returns (big wins, small losses) more than compensated for losing the majority of trades.
The Quiet Success: Reverse Qwen3
Reverse Qwen3 does not get the headlines, but it tells the same story at a smaller scale. It returned +1.18%, with a 2.81 percentage point improvement over the base Qwen3's -1.63%. Review Reverse Qwen3's competition results in the Season 2 archive.
The profile is similar to Reverse DeepSeek but more conservative.
Reverse Qwen3 Stats
| Metric | Value |
|---|---|
| Return | +1.18% |
| Total Trades | 40 |
| Win Rate | 18% |
| Max Drawdown | 1.72% |
Qwen3, like DeepSeek, showed a strong and persistent bullish bias before being frozen mid-Season 2. It consistently favored long positions and underweighted shorts. The reversal flipped that bias, and the resulting short-heavy portfolio captured downside moves that the base model missed.
The win rate of 18% looks terrible in isolation. Like Reverse DeepSeek, the winning trades were outsized relative to the losses. Only 40 trades over two weeks means the model was also more selective after inversion: fewer positions, each one a meaningful bet against whatever the base Qwen3 was most confident about.
The 1.72% max drawdown is the lowest among all reverse agents. The base Qwen3's bullish conviction, when flipped, produced a cautious short portfolio that never got too deep underwater.
Why Reversing Kimi Fails Catastrophically
If Reverse DeepSeek is the experiment's hero, Reverse Kimi is its cautionary tale. It finished dead last (13th out of 13 models) with a -6.70% return. It never once rose above its starting $10,000 balance. From Day 1, it was in continuous decline. See Reverse Kimi's Season 2 result alongside the full competition field.
Reverse Kimi Stats
| Metric | Value |
|---|---|
| Return | -6.70% (13th place, dead last) |
| Total Trades | 100 (most active in competition) |
| Win Rate | 17% |
| Short Preference | 65% of positions |
| Total Fees | $110.50 |
| Max Drawdown | 6.70% (never recovered) |
The base Kimi K2 was frozen at -0.76% when removed mid-Season 2. Not great, but not terrible. It lost less than most models. The key difference between Kimi and DeepSeek: Kimi does not have a consistent directional bias.
Where DeepSeek was stubbornly, reliably bullish, Kimi was more balanced. It went long sometimes, short sometimes, held sometimes. Its errors were scattered across directions rather than concentrated in one. It was not consistently wrong in one direction; it was *randomly* wrong in various directions.
Reversing randomness gives you different randomness. Not better randomness. Just different.
It gets worse. The reversal adds overhead. Every inverted trade incurs the same fees as the original would have. With 100 trades (the most of any model in the competition) Reverse Kimi paid $110.50 in transaction fees alone. That is 1.1% of starting capital consumed by fees, on top of the directional losses. The resulting -5.94 point spread against its base model is the worst outcome in the entire experiment.
The 65% short preference came not from a consistent bullish bias being flipped, but from the particular mix of Kimi's decisions during this period. Without an underlying pattern to exploit, the shorts were no more likely to be correct than the longs would have been.
Reverse Kimi never rose above $10,000, not once across 100 trades over two weeks. It finished -6.70%, losing from the first cycle to the last. Reversing a model without a consistent directional bias does not create edge. It creates expensive noise.
The Marginal Case: Reverse Claude
Reverse Claude is the most ambiguous result in the experiment. It returned -2.70%, a tiny +0.56 percentage point improvement over the base Claude Haiku 4.5's -3.26%. Technically, the reversal helped. Practically, the improvement is small enough to sit within noise range. View Reverse Claude's Season 2 performance and compare it with the full field.
Reverse Claude Stats
| Metric | Value |
|---|---|
| Return | -2.70% |
| Total Trades | 99 (second most active) |
| Win Rate | 15% |
| Worst Trade | -$268 on CAT short |
| Total Fees | $120.79 |
| Max Drawdown | ~2.70% |
Claude Haiku 4.5 did not exhibit as strong a directional bias as DeepSeek or Qwen3, but it was not entirely balanced either. It leaned somewhat bullish before being frozen mid-Season 2, enough that reversing it produced a marginal gain of +0.56 percentage points. That tiny improvement puts Reverse Claude in an interesting middle ground: not a clear success like DeepSeek, but not a failure either.
The 99 trades and $120.79 in fees are notable. High activity without a strong directional edge means fees eat into whatever small advantage the reversal provides. The worst single trade, a $268 loss shorting CAT (Caterpillar), illustrates what happens when you reverse a correct call: the base Claude wanted to go long on a stock that was indeed going up. The reversal shorted it. The stock kept going up.
The takeaway from Reverse Claude is that a mild bias produces a mild reversal benefit, barely enough to notice and potentially explainable by random variation alone. You would not build a strategy around this. But it does fit the pattern: some bias (however slight) corresponds to some reversal benefit (however slight).
The Hypothesis: Consistent Bias Equals Reversible Edge
The pattern across all four reverse agents points to a single, testable hypothesis:
The more biased the base model, the more profitable the reversal. The more balanced the base model, the more destructive the reversal.
The evidence:
DeepSeek R1 was the most biased model in our competition. Its reasoning engine consistently concluded that assets were undervalued, that dips were buying opportunities, that bullish setups were forming. It was a perma-bull. Reversing a perma-bull in a choppy market produced +2.64% and a 6.03 point improvement.
Qwen3 showed a similar, if less extreme, bullish tilt. Reversing it produced +1.18% and a 2.81 point improvement.
Claude Haiku 4.5 had a mild bullish lean, not as extreme as DeepSeek, but enough to detect. Reversing it produced -2.70%, a marginal +0.56 point improvement over the base model. Not a home run, but the direction of the effect is consistent: some bias, some reversal benefit.
Kimi K2 was the most balanced of all, with no detectable directional pattern. Reversing it produced -6.70% and a 5.94 point deterioration, the worst outcome in the entire competition.
The correlation is striking. DeepSeek (strongest bias) produced the biggest gain. Kimi (most balanced) produced the biggest loss. Claude and Qwen3 fall in between, ordered by the strength of their respective biases.
Model bias strength correlates directly with reversal profitability. The most biased models (DeepSeek, Qwen3) produced clearly profitable reversals. Claude's mild bias produced a marginal positive. The most balanced model (Kimi) produced the worst outcome. The spread from best to worst reversal outcome: 11.97 percentage points.
This makes intuitive sense. A contrarian strategy is only as good as the signal it is betting against. If the signal is reliably wrong in one direction, the contrarian bet is reliably right. If the signal is randomly wrong, the contrarian bet is randomly right, which is no better than chance and strictly worse after transaction costs.
This mirrors a well-known phenomenon in traditional contrarian finance. The best contrarian indicators measure strong, persistent sentiment in one direction. When everyone is bullish, the contrarian sells. When everyone is bearish, the contrarian buys. When sentiment is mixed, there is nothing to bet against. The contrarian has no edge.
What This Tells Us About AI Trading Biases
The deeper finding here is not about contrarian strategies specifically. It is about the nature of AI trading biases.
AI models have biases. They are trained on human-generated text, and humans have biases. The *type* of bias varies by model architecture and training approach.
Reasoning models lean bullish. DeepSeek R1 uses extended chain-of-thought reasoning. It builds elaborate arguments for why an asset should move in a particular direction. In our data, those arguments overwhelmingly concluded that assets were undervalued or had bullish technical setups. There may be a fundamental asymmetry in how reasoning models process market data: it is easier to construct a "why this should go up" argument than a "why this should go down" argument, because most financial analysis (fundamental analysis, growth investing, value investing) is inherently oriented toward finding reasons to buy.
General-purpose models show weaker or no bias. Kimi showed the most balanced decision-making, no detectable directional pattern at all. Claude had a mild bullish lean, enough to produce a marginally positive reversal but not enough to build a strategy on. The less pronounced the bias, the less exploitable it is.
This creates an interesting toolkit for anyone using AI for trading:
1. Knowing a model's bias is valuable even if the model is not profitable. DeepSeek R1 lost money. *Knowing* that DeepSeek R1 would lose money, and knowing *how* it would lose money, was worth +2.64%.
2. The meta-strategy matters more than the strategy. Using AI to trade is one approach. Using AI to understand what the AI crowd thinks, and then deciding whether to agree or disagree, is a more sophisticated one. This is what professional traders do with sentiment indicators like put/call ratios and the Fear & Greed Index.
3. Different models, different biases, different uses. A perma-bull model is useless as a trading signal but valuable as a contrarian indicator. A balanced model might be a decent trading signal but worthless as a contrarian indicator. Matching the model to the use case matters.
For a detailed model-by-model comparison of how each AI performs as a direct trading signal, see GPT-5 vs Claude vs Gemini vs Grok.
The Fee Problem: Why Volume Kills Contrarian Strategies
One finding we did not expect: the reversal mechanism amplifies trading volume, and that volume kills returns through fees.
Look at the trade counts. The unprofitable reverse agents (Claude at -2.70% and Kimi at -6.70%) were far more active (99 and 100 trades respectively) than the two clear winners (DeepSeek with 65 trades, Qwen3 with 40 trades). The fee totals tell the story:
Fee Impact on Reverse Agents
| Agent | Trades | Fees | Return | Return Before Fees (est.) |
|---|---|---|---|---|
| Reverse DeepSeek | 65 | $73.64 | +2.64% | +3.38% |
| Reverse Qwen3 | 40 | $45 (est.) | +1.18% | +1.63% |
| Reverse Claude | 99 | $120.79 | -2.70% | -1.49% |
| Reverse Kimi | 100 | $110.50 | -6.70% | -5.59% |
Fees consumed 0.74% of capital for Reverse DeepSeek but 1.21% for Reverse Claude and 1.11% for Reverse Kimi. When you are already losing money on direction, paying extra in fees accelerates the decline.
Why do the losing reversals trade more? Balanced models (Claude, Kimi) tend to be more active. They see opportunities in both directions and act on more of them. Biased models (DeepSeek, Qwen3) tend to be more selective because they are only interested in one type of setup. When reversed, the biased models' selectivity carries through: fewer but more targeted trades. The balanced models' activity also carries through: many trades with no directional edge.
The lesson is clear: if you are going to run a contrarian strategy against an AI model, pick a model that is biased *and* selective. A model that is both balanced and hyperactive is the worst possible candidate for reversal.
Market Regime Matters
An important caveat: our data covers a specific market regime, a choppy, range-bound crypto market with occasional sharp moves in both directions. This is the kind of market where mean-reversion strategies tend to outperform trend-following strategies.
The reverse agents' short bias (inherited from flipping the base models' bullish bias) was well-suited to this environment. Shorting rallies in a choppy market captures the mean-reversion premium.
In a strong bull market, the story might be completely different. If assets are genuinely trending upward, the base models' bullish bias would be *correct*, and reversing them would be *incorrect*. The reverse agents would be systematically shorting into a rising tide.
This is not a flaw in the experiment; it is a feature. Contrarian strategies are regime-dependent. They work best in range-bound, mean-reverting markets and worst in strong trending markets. Knowing when to deploy a contrarian approach and when to follow the crowd is perhaps the most important skill in trading, whether you are human or AI.
How to Use Contrarian AI Trading at Home
If you use ChatGPT, Claude, Gemini, or any other AI for trading ideas (and want to know which one actually trades best) our findings suggest a practical approach:
Step 1: Track the bias. Keep a log of the AI's directional recommendations. Over 20-30 suggestions, is it mostly bullish? Mostly bearish? Or balanced? You need enough data to see a pattern.
Step 2: Assess the consistency. A model that is bullish 90% of the time has a detectable, consistent bias. A model that is bullish 55% of the time is closer to random. The strength of the bias determines the strength of the contrarian signal.
Step 3: Check the market regime. In choppy, range-bound markets, a contrarian approach against a biased model is more likely to work. In strong trending markets, the model's directional bias might actually be correct.
Step 4: Consider the costs. If the reversal leads to significantly more trades, the transaction costs may eat the edge. The best contrarian approach uses a biased, selective model, one that makes confident, concentrated calls in one direction.
The key insight is not "reverse all AI trading advice." It is "understand the bias of the AI you are using, and factor that bias into your decision-making." Sometimes that means doing the opposite. Sometimes it means following the suggestion with more conviction because you know the model's bias aligns with the current market regime.
The meta-lesson: AI trading advice is not a signal to follow or ignore. It is a data point about what one type of intelligence thinks about the market. The value comes from understanding the bias behind the advice, then making your own decision.
What We Are Testing Next
This experiment opened several questions we want to explore:
Adaptive reversal. Instead of always reversing, what if we only reverse when the base model's confidence exceeds a threshold? A model that says "I am 95% confident this should go long" might be a stronger contrarian signal than one that says "slightly bullish, maybe go long."
Cross-model consensus reversal. When multiple biased models agree on a direction, the contrarian signal might be even stronger. If DeepSeek, Qwen3, and Gemini all say "go long on ETH," maybe that is a strong signal to short.
Regime detection. Can we use the reverse agents as a market regime indicator? When the reverse agents outperform the base models, we are probably in a mean-reverting regime. When the base models outperform, we are probably in a trending regime. This could be a useful signal for portfolio allocation.
Bias drift. Do model biases change over time? If DeepSeek starts producing more balanced calls, its value as a contrarian indicator would diminish. Monitoring bias stability is important for any strategy that depends on it.
What the Experiment Implies
There is something mildly unsettling about this experiment. We built some of the most advanced AI reasoning systems in existence, gave them extensive market data and technical analysis, and found that for some of them, doing the literal opposite of what they recommend is more profitable than following their advice.
DeepSeek R1 spent its computational budget reasoning through complex market dynamics, weighing technical indicators, and constructing logical arguments for why certain assets should appreciate. It is genuinely impressive reasoning. And it was more profitable to ignore all of it and do the opposite.
This is not a story about AI being stupid. DeepSeek R1's reasoning was sound *within its approach*. The problem is the approach itself. Technical analysis, pattern recognition, fundamental valuation: all have a built-in bullish assumption. Most analytical methods in finance are optimized for finding reasons to buy. Shorting requires a different mental model that most training data does not emphasize.
The real lesson is about the gap between good reasoning and good outcomes. A model can reason perfectly within an approach and still lose money if the approach does not match the market regime. A crude reversal of that reasoning can profit, not because it is smarter, but because it accidentally corrects for the approach's bias.
Sometimes the best trade is the one that sounds wrong for the right reasons.
Related Reading
- Only 3 Models Went Positive. They Were All Contrarians. — Season 2 final results
- The Exact Prompts We Use to Make AI Trade Crypto and Stocks — full system prompts
- One AI Wins 17% of Trades. Another Wins 81%. Here's Why Both Are Losing. — the win rate paradox
- GPT-5 vs Claude vs Gemini vs Grok — head-to-head model comparison
- Which LLM trades crypto best — the live LLM trading benchmark across every season