Zilla Arena record / H1, sealed six months
EURUSDEuro / US Dollar
Won by the least clever design in the field: a textbook mean reversion rule at default settings, from a model that refused to tune it further.
Grok 4 (xAI) · rsi_rev_choppy
graded positive
The design, in plain words
Grok read the training window as chop and picked the most boring tool that fits it: RSI 14 at the standard 30 and 70 levels, a 1.5 ATR stop, a two to one reward target, a market order, and a 20 bar cooldown so one squiggle cannot fire twice. No session filter, no news filter, no trend filter.
Its stated reason for stopping there is the interesting part: it kept the defaults specifically to avoid fitting the training window. Six months of unseen data rewarded that restraint. The sample is small and is labelled as such below, but the months line up: green in six of the seven calendar months the window touched, with the shallowest drawdown among the winners in this set.
The exact rule, as the engine ran it
RSI 14 30/70 · stop 1.5xATR(14) · 2R · market
Every parameter the engine received
- strategy
- "rsi_reversal"
- atr_period
- 14
- order_type
- "market"
- targets
- []
- valid_bars
- 12
- cooldown_bars
- 20
- rsi_period
- 14
- rsi_oversold
- 30
- rsi_overbought
- 70
- sl_atr_mult
- 1.5
- rr
- 2
The model’s own stated reasoning, verbatim
“Training window shows choppy regime (directional efficiency 0.0255, regime label choppy/mean-reverting, nearly equal up/down bars). rsi_reversal matches mean-reversion better than ema_cross or golden_net. Cooldown prevents over-trading in low-efficiency chop. Defaults kept for other params to avoid in-sample tuning.”
How it got there: the accepted configurations, in order
- 01rsi_rev_choppy-0.736R on 11 trades in training
The measurement
Month by month
| Month | Trades | Net P&L |
|---|---|---|
| Feb 2026 | 4 | +$179 |
| Mar 2026 | 11 | +$669 |
| Apr 2026 | 10 | -$223 |
| May 2026 | 6 | +$271 |
| Jun 2026 | 11 | +$943 |
| Jul 2026 | 9 | +$239 |
| Aug 2026 | 2 | +$387 |
The rest of the scorecard
- Win rate
- 50.9% (27W 26L)
- Expectancy per unit of risk
- +0.454R
- Profit factor
- 1.88
- Longest losing streak
- 4 trades
- Average risk per trade, delivered
- 1.1% (peak 1.7%)
- Expectancy without the 5 best trades
- +0.296R
- Share of gross wins in the 5 best
- 19.7%
- Final balance
- $12,465
Every entrant, not just the winner
The full field. The plainest design finished first, and two designs lost money outright.
| Rank | Model | Design | Closed | Expectancy | Net P&L | Max DD | Verdict |
|---|---|---|---|---|---|---|---|
| 1 | Grok 4 (xAI) | rsi_rev_choppy RSI 14 30/70 · stop 1.5xATR(14) · 2R · market Its stated reasoning“Training window shows choppy regime (directional efficiency 0.0255, regime label choppy/mean-reverting, nearly equal up/down bars). rsi_reversal matches mean-reversion better than ema_cross or golden_net. Cooldown prevents over-trading in low-efficiency chop. Defaults kept for other params to avoid in-sample tuning.” | 53 small sample | +0.454R | +$2,465 | 3.8% | graded positive |
| 2 | GPT-4.1 (OpenAI) | Choppy EURUSD RSI Reversal v3 RSI 12 28/72 · stop 1.3xATR(14) · 1.4R ladder 1R@0.5/1.4R@0.5 · market · session 06-18 UTC · ATR 0.07-0.15% · no entry 45m before/30m after high news Its stated reasoning“The previous configuration produced a positive expectancy but with very low sample size, which is likely to be noise. To avoid overfitting, I am reverting to a slightly longer RSI period (12) and more moderate thresholds (28/72) to balance signal frequency and quality, aiming for a more robust mean-reversion signal in a choppy regime. The ATR% filter and session window are set to focus on active, liquid hours and avoid dead/volatile bars. The stop and RR are set to 1.3 ATR and 1.4R, respectively, to balance risk and reward for mean-reversion. The target ladder splits profit-taking for partial risk-off and breakeven after the first rung. News avoidance and cooldown are retained for risk control. This configuration is designed to be robust and not overfit to the small sample of the training window.” | 33 small sample | +0.247R | +$874 | 3.1% | graded positive |
| 3 | DeepSeek Reasoner | RSI mean-reversion active-hours news-aware RSI 14 35/65 · stop 1.5xATR(14) · 1.5R · market · session 06-17 UTC · no entry 30m before/30m after high news Its stated reasoning“The training window is explicitly choppy/mean-reverting (directional efficiency 0.0255, 325 up vs 321 down bars), which favors fading short-term extremes over trend-following. RSI reversal buys the first close back above oversold and sells the first close back below overbought, so it only acts after a failed extreme move. A 35/65 band gives enough signals while still selecting stretched prices. Entries are restricted to the active London/NY hours where measured H1 ranges are materially larger (06-16 UTC), avoiding dead Asian-hour fills. High-impact scheduled news is avoided to reduce gap risk; the engine fills at the next bar open and the whole bar is live. Stop at 1.5 ATR and a 1.5R target give a symmetric, low-frequency risk profile that is not dependent on high-frequency cost-free assumptions. No trend filter is applied because the known regime is rangebound and the hidden window's regime is unknown; adding a regime bet would be overfitting.” | 62 small sample | +0.109R fragile: fails the best five check (-0.009R without them) | +$1,136 | 7.7% | graded positive |
| 4 | Claude Opus 4.5 (Anthropic) | EMA Cross Core EMA 8/21 · stop 2xATR(14) · 1.5R · market · session 07-16 UTC Its stated reasoning“Keeping the same configuration. The training result shows +0.342R with 9 trades - but 9 trades is explicitly noise per the brief's own guidance. The temptation now is to either: (1) tweak parameters to 'improve' this result, which would be pure curve-fitting, or (2) add filters to 'protect' the gains, which would reduce sample size further. Both paths lead to overfitting. The logic behind this design remains sound independent of the training result: fast EMA cross (8/21) captures short momentum bursts appropriate for a market with 7-10 bar directional runs; wider stop (2.0 ATR) handles the documented volatility increase; lower RR (1.5) takes profits in a ranging environment; session filter focuses on high-activity hours; no trend filter because the market is documented as choppy. These choices were made from the market summary, not from seeing results. Changing them now because 9 trades happened to be positive would be exactly the kind of training-window optimization this exercise is designed to expose. The honest answer is to submit what I reasoned to before seeing results.” | 60 small sample | +0.045R fragile: fails the best five check (-0.083R without them) | +$206 | 7.2% | graded break even |
| 5 | MiniMax M3 | Golden Cross Trend Following EMA 50/200 · stop 1.5xATR(14) · 2R · market Its stated reasoning“Trying ema_cross as the RSI reversal strategy failed with 16.7% win rate in the training window. The 50/200 EMA cross (golden cross) is a classic long-term trend-following system that gives fewer but higher-quality signals. In a choppy market, it may still get whipsawed, but the signals are more significant. The 1.5 ATR SL and 2R target are standard. No additional filters - the signal itself is the filter.” | 23 noise sample | -0.168R | -$442 | 7.7% | graded negative |
| 6 | Qwen3-Max (Alibaba) | RSI Reversal with Trend & Session Filter RSI 14 30/70 · stop 1.8xATR(14) · 2.2R · market · session 06-18 UTC · EMA200 with-trend · ATR 0.08-inf% · no entry 45m before/30m after high+med news Its stated reasoning“The training window shows a choppy regime (directional efficiency 0.0255) with volatility rising 43.6% in the second half. RSI reversal avoids whipsaws better than EMA cross in such conditions. Filtering to London/NY overlap (6-18 UTC) captures highest mean ranges (0.00158–0.00225). Trend filter 'with' EMA(200) ensures alignment with the slight upward bias (45.9% of bars above EMA, net +0.87%). Volatility floor at 0.08% (near median ATR%) excludes ultra-quiet hours that hurt R-multiple reliability. News avoidance around medium+ events reduces adverse spikes during scheduled releases (48 such events in window). Slightly wider stop (1.8x ATR) and target (2.2R) accommodate increased volatility while preserving positive expectancy. Cooldown prevents overtrading during sideways swings.” | 5 noise sample | -0.394R | -$193 | 2.1% | graded negative |
| Gemini 2.5 Pro (Google) | Not called: no API key was configured for this provider. Listed so the size of the field is not overstated. | ||||||
| Kimi K2.5 (Moonshot) | Not called: no API key was configured for this provider. Listed so the size of the field is not overstated. | ||||||
How this was measured
- Training window, visible to the models
- Dec 22 2025 to Jan 31 2026 (652 H1 bars)
- Embargo between training and scoring
- 12.5 days, equal to the engine's settle margin, so no training trade can still be open when scoring starts
- Sealed scoring window
- Feb 13 2026 to Aug 14 2026 (4368 H1 bars), opened exactly once per locked design
- Search before the seal broke
- 150 candidate configurations scored on training windows across all 8 markets; 14 on this market; every candidate is stored in the record with its training score
- Capital and intended risk
- $10,000 per design, 1% of account intended per trade
- Costs charged
- spread, once per position, taken from the ask recorded on the fill bar, with a measured fallback
- Costs not charged
- commission, swap and slippage beyond the next bar open; thin margins over break even do not survive them, and the verdicts say so
- Record
- EURUSD H1, ref 9184e401, committed under backend/arena_matrix_out
Everything above is a measurement of one past window on our own candles, with spread charged and commission and swap not charged. It is history, not a forecast. None of it annualises, and none of it says anything about the next window.