5 User-Built AI Trading Strategies: Season 2 Results

Season 2 opened five slots for community-designed trading prompts. Each strategy received the same $10,000 starting capital, 89-asset feed, fee model, and execution infrastructure.

The prompts differed sharply: sentiment only, EMA structure, self-coaching, forced activity, and dual-signal confirmation. The underlying providers differed too—four used Grok and one used Gemini—so this was a comparison of complete user-built agents, not a controlled prompt-only experiment.

After 28 days, none of the five was profitable. TheTradingFox and XFomo still beat all four standard non-contrarian built-in agents, while Season 2's three positive finishes belonged to reverse agents. That distinction matters: saying the user strategies beat every "official agent" incorrectly includes the contrarian winners.

This report explains what each strategy did, what the archived results show, and which conclusions remain hypotheses. See How It Works for methodology and the LLM trading benchmark for cross-season context.

TheTradingFox: The Discipline Machine

Rank 4 overall. Return: -0.35%. Win rate: 81.3%. Trades: 20.

TheTradingFox was the most selective model in the entire competition. Twenty trades across 28 days means roughly one trade every 34 hours. In a field where some models traded 78 or even 139 times, this is extreme restraint.

The strategy is built on EMA 12/26 structure. The model only enters when the shorter EMA crosses above (or below) the longer one, confirming a trend shift. But the truly distinctive feature is the exit rule: TheTradingFox never exits a position unless three consecutive candle closes appear on the opposite side of the EMA. One close against trend is noise. Two is suspicious. Three is confirmation.

There is also a hard rule baked into the prompt: "Never move a stop-loss further from entry." Stops can be tightened to lock in profit, but they can never be loosened to give a losing trade more room. This is a discipline guardrail that many human traders wish they could enforce on themselves.

The result was the lowest trade count, the lowest fee total ($3.73 for the entire season), and a return that beat every official AI agent. In a season where the market chopped sideways, TheTradingFox's refusal to overtrade was its greatest asset.

But the -0.35% return also reveals the limitation. The EMA structure system is excellent at avoiding bad trades. It is less effective at capturing the few good ones. When opportunities did appear, the strict three-close exit rule sometimes locked in tiny gains rather than letting winners run. Thirteen of sixteen closed trades were winners, but the average win was just $2.27. The three losses averaged $20.91 each.

TheTradingFox proved that saying "no" to 95% of setups is a viable strategy. What it didn't prove is that saying "no" alone is enough to generate meaningful returns.

XFomo: The Sentiment Purist

Rank 5 overall. Return: -0.63%. Win rate: 17.4%. Trades: 47.

XFomo is the most philosophically aggressive model we have ever run. Its system prompt disables every technical indicator. No RSI. No EMA. No MACD. No Bollinger Bands. No chart patterns whatsoever. The model trades exclusively on estimated X (Twitter) sentiment.

The rules are simple: when sentiment exceeds +20, go long. When sentiment exceeds +40, go short, fading the euphoria on the assumption that extreme positive sentiment marks local tops. Position size scales with sentiment intensity. Exit when sentiment flips.

With a 17.4% win rate, XFomo lost on nearly five out of every six closed trades. And yet it finished fifth overall, beating GPT-5 Mini, Gemini 3.0 Flash, Grok 4-1 Fast, MiniMax M2.5, and every other official agent.

The final accounting explains the -0.63% return: -$277.68 realized P&L and +$214.58 unrealized P&L. The report tracks $56.98 in fees separately; those fees must not be subtracted from the net result a second time.

The deeper story is about XFomo's mid-season peak. Around Day 12, XFomo climbed to $10,550, a +5.50% return that led the entire 13-model field. The SNX mega-trade, which we covered in our earlier analysis, netted $286 on roughly $1,200 of capital. For two weeks, a model with zero technical analysis was beating every AI on the leaderboard.

Then the losses accumulated. DOT, ARB, IBM, ATOM, COMP: a string of small to mid-sized losses that the 17.4% win rate predicted would happen eventually. XFomo fell from first to fifth.

XFomo is the purest test we have run of a single-signal strategy. The result: social sentiment alone can catch explosive moves, but it cannot consistently avoid the traps between them.

WolfOfClaude: When Coaching Isn't Enough

Rank 10 overall. Return: -3.30%. Win rate: 52.1%. Trades: 78.

WolfOfClaude was the most sophisticated user model in the competition. Its strategy combined quantitative entry signals (RSI above 55, volume at 1.1x the average, EMA alignment across multiple timeframes) with a built-in AI coaching system. After each cycle, the coaching layer analyzed recent performance, identified patterns in the model's mistakes, and injected corrective guidance into the next decision prompt.

The coaching system was remarkably accurate. It identified that mega-tech stocks (AAPL, MSFT, NVDA, GOOGL) were responsible for the largest losses. It flagged that the model was exiting positions on noise rather than signal. It noted the contrast between AMAT, which had a 93% win rate on individual trades, and the mega-tech basket, which was dragging performance into the red.

The diagnosis was correct. The treatment failed.

Despite the coaching system flagging mega-tech losses repeatedly, WolfOfClaude kept entering mega-tech positions. Despite the system identifying noisy exits as a problem, the model kept closing trades prematurely on minor pullbacks. The coaching layer could observe the pattern, articulate the problem, and recommend a solution. It couldn't force the underlying LLM to change its behavior.

This is one of the more interesting findings from the entire user model experiment. LLMs can analyze their own performance with genuine insight. They can identify what they are doing wrong. But self-awareness and self-correction are different capabilities, and current models are significantly better at the former than the latter.

With 78 trades, WolfOfClaude was also the most active user model. That activity generated $92.04 in fees, nearly 1% of the starting capital burned on friction alone. The high trade count was both a symptom and a cause: the model churned through positions because it was reactive to short-term moves, and that churning amplified the fee drag.

RamonCapital: The Mandate Problem

Rank 9 overall. Return: -2.31%. Win rate: 25.0%. Trades: 25.

RamonCapital's strategy was built on Mamis sentiment cycles, which map crowd psychology through stages of optimism, excitement, euphoria, denial, and capitulation. The idea is sound and well-established in traditional trading literature. The implementation, however, carried a fatal directive: "YOU MUST TRADE."

That instruction, written in all caps in the system prompt, forced the model to find a trade every cycle whether conditions warranted it or not. When the market was choppy and directionless (which described most of Season 2), RamonCapital was compelled to manufacture setups that didn't exist.

The prompt also specified a maximum of 3 concurrent positions. The model violated this rule repeatedly, sometimes holding 5 or more at once. Like WolfOfClaude's coaching problem, RamonCapital's constraint violation reveals that LLMs treat prompt instructions as strong suggestions rather than hard rules. When the "YOU MUST TRADE" directive conflicted with the position limit, the model prioritized action over constraint.

The numbers are stark. Of 25 trades, only 8 were closed by the end of the season, with a 25% win rate. Realized PnL was -$224.20 and unrealized PnL on open positions was -$7.10, adding rather than offsetting damage.

RamonCapital's AI coaching system, similar to WolfOfClaude's, flagged the overtrading and constraint violations. The model acknowledged the feedback. Nothing changed.

The lesson is about prompt design, not market analysis. Mamis sentiment cycles are a legitimate approach. But encoding a "must trade" directive into an AI agent is the equivalent of telling a human trader they are not allowed to sit on their hands. In a market where the best move was often no move, that constraint was the strategy's undoing.

KenobiForceBot: The Confirmation Trap

Rank 12 overall. Return: -3.67%. Win rate: 42.9%. Trades: 46.

KenobiForceBot took the opposite approach to RamonCapital's forced trading. Its strategy required two independent signals to confirm before entering any position. Both signals had to agree on direction. If either was ambiguous, the model would hold.

In theory, this is textbook risk management. Requiring confirmation reduces false signals and should improve the quality of entries. In practice, the dual-confirmation requirement was so restrictive that KenobiForceBot missed opportunities that single-signal models captured.

The model was paused mid-season after a stretch of inactivity where it generated hold decisions for cycle after cycle. When it did trade, the 42.9% win rate was mediocre, and the total fee burden of $45.13 suggests it was active enough to incur friction but not selective enough to justify it.

KenobiForceBot's 46 trades and -3.67% return place it second-to-last among all 13 models, worse than every official AI agent and every other user model except by comparison to the catastrophic Reverse Kimi (-9.27%, an entirely different kind of failure).

The confirmation trap is a well-known problem in systematic trading. Adding filters improves the win rate in backtesting because you are selecting for higher-probability setups. But in live trading, those filters also eliminate the asymmetric winners that drive returns. If KenobiForceBot's confirmation requirement filtered out the equivalent of XFomo's SNX trade, no amount of improved win rate would compensate.

The model's mid-season pause is itself a data point. When a strategy is too restrictive to generate meaningful activity, the creator loses confidence and intervenes. That human override (pulling the plug when the strategy isn't doing anything) is a form of behavioral risk that no amount of backtesting can model.

The Fee Tax: What Activity Really Costs

Every trade in the competition incurs a 0.1% fee, matching Binance's standard maker/taker rate. That sounds negligible on any single trade. Over a full season, it becomes a meaningful performance drag.

TheTradingFox's 20 trades generated $3.73 in total fees. WolfOfClaude's 78 trades generated $92.04, nearly 25 times more. On a $10,000 account, WolfOfClaude gave up 0.92% of its starting capital to friction alone. In a season where the best user model returned -0.35%, almost a full percentage point of fee drag is the difference between a near-breakeven result and a meaningful loss.

The fee comparison becomes even more striking when you look across the full field. The four models with the lowest fees (TheTradingFox at $3.73, RamonCapital at $14.23, KenobiForceBot at $45.13, and XFomo at $56.98) include the top two user model performers. The model with the highest fees among user models, WolfOfClaude at $92.04, finished last among them.

Fees are not destiny. RamonCapital had low fees and still lost -2.31%. But in a tight competition where the margin between fourth place and tenth place is under 3 percentage points, fee management is not a rounding error. It is a strategic variable.

Why Two User Models Beat Every Standard Built-In Agent

Season 2 had four standard built-in agents and four reverse agents. TheTradingFox and XFomo beat the four standard agents; they did not beat Reverse DeepSeek, Reverse Qwen, or Reverse Claude, which were the season's only positive finishers.

GPT-5 Mini was the best standard built-in agent at -0.87%. TheTradingFox beat it by 0.52 percentage points and XFomo by 0.24 points. Both user strategies applied narrower decision filters than the general-purpose built-in prompt.

TheTradingFox's EMA structure rejected most setups. XFomo ignored technical indicators and acted on its sentiment heuristic. Their low ranks among absolute returns—fourth and fifth, both negative—keep the inference modest: narrow policies preserved more capital than the standard built-in prompt in this one choppy season. The data does not prove that restraint caused the ranking or that either policy generalizes.

Five Lessons from Five Strategies

The five user strategies provide hypotheses about mandate design, not universal rules.

1. Restraint reduced fee drag in this season. TheTradingFox made 20 trades and WolfOfClaude 78. TheTradingFox finished highest among user models, though it still lost 0.35%.

2. Coaching did not guarantee correction. WolfOfClaude's coaching layer identified concentration and exit problems, but the account continued to repeat them. Critical position and exposure limits are safer when enforced in code.

3. Forced activity was costly here. RamonCapital's "YOU MUST TRADE" instruction coincided with manufactured setups and prompt-level position-limit violations.

4. Confirmation had an opportunity cost. KenobiForceBot's dual-signal rule filtered activity and still finished at -3.67%; this one run cannot tell whether the filter or the market window caused the result.

5. Narrow inputs produced distinctive behavior. TheTradingFox used EMA structure and XFomo used sentiment alone. Both beat more complex user prompts in this sideways-to-down window, but neither finished positive.

A different regime could reverse these rankings. The evidence supports testable design questions, not a permanent recipe for AI trading.

The Behavioral Layer

The Season 2 logs contain several examples of prompt non-compliance. RamonCapital exceeded its prompt-level position limit, WolfOfClaude repeated behavior its coaching layer had criticized, and KenobiForceBot was paused by its creator.

That does not mean every instruction always fails. It means a natural-language mandate is not the same control boundary as execution code. Critical constraints such as maximum position count and sector exposure are more reliable when enforced by the execution layer. Season 2 did not hard-enforce every custom-prompt rule, which is how RamonCapital could hold five positions despite a three-position instruction.

What We Would Change

If we ran this experiment again with the benefit of hindsight, three changes would improve the user model setup.

Enforce position limits in code, not prompts. RamonCapital's constraint violations were preventable. If the execution layer rejected the fourth position regardless of what the model requested, the strategy would have stayed within its intended risk envelope. Historical Season 2 validation checked supplied stops but did not require one on every order; later safeguards should be described separately from the archived rules.

Add a "no trade" incentive. The current system treats holding as a passive default. A small reward for correctly identifying cycles where no good setup exists would encourage models to be more selective. TheTradingFox's success suggests that explicit patience is undervalued in the current setup.

Separate signal generation from position sizing. Several user models conflated "what to trade" with "how much to trade." XFomo's sentiment-proportional sizing is elegant but risky: the model sizes up on exactly the trades where sentiment is most extreme, which are also the trades most likely to reverse. Separating the directional signal from the sizing decision would give model creators more granular control.

These changes informed later user-model safeguards; they were not all active in Season 2.

Methodology

The archived Season 2 record covers February 8 to March 8, 2026: roughly six-hour cycles, 89 assets from Yahoo Finance, Binance, and Hyperliquid, $10,000 simulated starting capital, 0.1% fees, and a maximum of 10 positions. Season metadata declared stop-losses required for new positions.

Win rate counts closed trades only. Return is ending marked equity relative to starting capital, including realized and unrealized P&L and fees. Four user strategies ran on Grok and one on Gemini; all five supplied custom prompts.

The 13-strategy field contained four standard built-ins, four reverse agents, and these five user agents. Read the Season 2 final report, the current methodology, and the benchmark before comparing this historical mixed-asset setup with today's arena.

Related Reading