Arena Blog

LLMs don't trade profitably. But why would anyone expect them to? The honest answer is uncomfortable: nobody can see the future, and markets are explicitly the opposite of a forecasting machine — they're a mechanism for the ongoing negotiation of disagreement. On the Grossman-Stiglitz paradox, why one side of every trade is convinced it's smarter than the other, and why AI trading tools mathematically neutralize themselves the moment they become widespread.

We wrote 500 lines of rules. Claude read them — and still wrote text-[13px] instead of text-xs. The mistake wasn't Claude. The mistake was confusing documentation with enforcement.

Grid bots are marketed as "low-risk" and, with AI, as a source of "consistent profits." We checked the claims against the mechanics: a grid earns on sideways oscillation, not trends — and "always trading" means trend-chasing. A fair breakdown, plus how to actually evaluate a grid bot.

AI agents pay per API call, per query, per compute slice — median $0.01–0.10. That kills card rails and makes the blockchain architecture the real question. A detailed comparison of Ethereum, Solana and Sui against the actual requirements (sub-cent fees, sub-second finality, programmable spend limits, identity) — plus the purpose-built payment chains Tempo and Arc that may be technically further ahead but barely adopted. The honest answer has no single winner.

x402 moves $28,000 in daily volume. Tron-USDT moves $20-30B. Difference: six orders of magnitude. Yet both get sold as "the future of stablecoin settlement." A sober reframing of why x402 is infrastructure-building and option value, while Tron-USDT is the shadow-dollar standard of the Global South — and why both are real without being the same market.

Anyone who has never set up a grid bot fails at the first step: range too tight → bot is immediately out of the corridor. Range too wide → barely any trades, no profit. We've built an auto mode into the grid backtest that suggests range and grid count based on 7-day volatility — KuCoin style. Plus the custom mode for power users that was already live.

A 5-asset what-if simulator: Bitcoin, S&P 500, Ethereum, Gold and Cash. See your mix vs. 100% Bitcoin — with Sharpe, max drawdown and volatility, not just end value.

A single number from 0 to 100, but built from 8 market indicators instead of one. Bullmarket stage, MVRV, Mayer, funding rates, hash ribbons — aggregated, transparent, free.

For 18 months we've been quietly building Backtesting Arena — a platform where 500+ users have run 10,000+ backtests across Bitcoin, stocks, ETFs, commodities, and forex. Daily cycle scores, on-chain indicators, sentiment dashboards, strategy insights. All powered by the same data layer that's been running on a private quasi-API.

Six frontier LLMs got $10,000 each and were told to autonomously trade crypto perps. Four lost more than 60 percent. The two winners were Chinese. What this really tells us isn't "China wins AI." It's: LLMs are language probability models, not market actors. A sober look at the latest empirical evidence that autonomous AI trading doesn't work.

CAGR assumptions are smooth. Bitcoin isn't. Anyone planning in a hypothetical world of constant 20% growth is planning for a world that never existed. We've added a Historical Replay mode to the Bitcoin Lifestyle Calculator — pick a start year and we run your strategy through the actual BTC history with Mt.Gox aftermath, COVID crash, FTX collapse, and ATH hype. Here's why that's often brutal and always honest.

Most BTC lifestyle calculators give you ONE number: "in 20 years you'll have X." That's fantasy, not planning. Bitcoin runs at ~70% annualized volatility — the expected value is just one of millions of possible stories. We've added two risk lenses to the Bitcoin Lifestyle Calculator: a Multi-Path Stresstest and a 1000-path Monte Carlo simulation. Here's what they tell you and why you need both.