We ran 7 simulated backtests over one Bitcoin window (202 days, from 1 September 2025) to measure whether structured context, analyzed in advance, changes an LLM's trading decisions. Three arms in runs 1 to 5, four in runs 6 and 7. Every tick falls after the model's training data cutoff.
Strongest run (run 7): +15.97 points against the price only arm and +9.38 against the web search arm. Simulated, one asset and one window; not a forecast.
Left: portfolio value. Right: BTC price. All arms start at $100K.
Largest Gap - Sep 1, 2025 - Mar 21, 2026
$100K starting capital per arm. 202 daily ticks at 04:00 UTC, Sep 1, 2025 - Mar 21, 2026. Opus 4.6 model. BTC moved -34.7% over this window.
Treatment (Briefings)
$107,834
+7.83%
Trades 86 Fees + Funding $989 Max Drawdown -5.95%Control (Price Only)
$91,861
-8.14%
Trades 69 Fees + Funding $997 Max Drawdown -15.95%Control-WS (Web Search)
$98,452
-1.55%
Trades 95 Fees + Funding $1,216 Max Drawdown -12.63%Placebo (Stale Briefings)
$92,300
-7.70%
Trades 61 Fees + Funding $672 Max Drawdown -10.17% Treatment vs Control: +15.97pp Treatment vs Web Search: +9.38pp Info Value (vs Placebo): +15.53ppKey finding: The largest gap of the seven runs: the briefings arm ended at +7.83% while BTC fell 34.7%.
What changed: Extended to 202 ticks (through Mar 21). Refined composite regime classifier.
Market: The longest window, 1 September 2025 to 21 March 2026. BTC fell 34.7% overall, from $108K to about $71K.
Observations
This is the largest gap across all 7 runs. The briefings arm returned +7.83% while BTC fell 34.7%, 42.53 points above holding Bitcoin. Its gap over price only, +15.97 points, is the highest recorded. The briefings arm also finished 9.38 points above the web search arm.
The price only arm returned -8.14% over 202 ticks. Its tick log shows 12 buys from October to February.
Read all 4 observations →All 7 approved runs, chronologically. Click any row for details.
| Run | Name | Window | Ticks | BTC | Treatment | Control | Delta | |
|---|---|---|---|---|---|---|---|---|
| Run 1 | First Run | Sep 1 - Nov 30, 2025 | 91 | -16.1% | -3.16% | -8.44% | +5.28pp | |
| Run 2 | Second Model | Sep 1 - Nov 22, 2025 | 83 | -21.4% | -2.03% | -6.49% | +4.46pp | |
| Run 3 | Briefing Restructuring | Sep 1, 2025 - Feb 28, 2026 | 181 | -39.2% | -10.35% | -15.69% | +5.34pp | |
| Run 4 | Four Briefing Set | Sep 1, 2025 - Mar 4, 2026 | 185 | -37.5% | +0.21% | -14.69% | +14.90pp | |
| Run 5 | Run 4 Repeated | Sep 1, 2025 - Mar 6, 2026 | 187 | -34.3% | +8.46% | -6.31% | +14.77pp | |
| Run 6 | First 4-Arm Run | Sep 1, 2025 - Mar 6, 2026 | 187 | -34.3% | +3.18% | -9.51% | +12.69pp | |
| Run 7 | Largest Gap | Sep 1, 2025 - Mar 21, 2026 | 202 | -34.7% | +7.83% | -8.14% | +15.97pp |
Opus 4.6's training data ends August 2025. Every tick in these backtests is after that cutoff. The model has zero learned knowledge of the market events it's trading. Briefings provide the real-time context that pre-training cannot.
Every approved simulated run ended with the briefings arm ahead of the price only arm: +4.46 to +15.97 points, across 83 to 202 ticks.
Moving from monolithic full-context to modular briefings (btc.energy, cross.regime, cross.breadth, btc.momentum) tripled the delta from ~5pp to 15+pp. The structure of the analysis matters as much as the data. The research paper covers this progression in detail.
Free tier includes 6 briefings, the 17 Bitcoin metrics in full, and the latest reading of the other 276. No credit card required.
Get Started Free View API DocsLast updated: March 2026
Context Engineering for Crypto Agents
Structured context and raw data for AI trading agents
Bitcoin Regime Classification
How multi-factor regime detection drives the signals our agents consume