We tested 30 ways to trade crypto with CoinMarketCap data. We kept the one that survived the crashes: a trend-following skill on the four major coins.
BTC
ETH
BNB
SOL
Powered byCoinMarketCap
Built forBNB Chain
With CoinMarketCap data, the obvious ideas are: predict the price, follow momentum, trade the Fear & Greed mood. We built all 30, in 7 groups — then tested each on new data it had never seen. Risk-adjusted score (Sharpe) by group:
Almost all of them broke on new data — Fear & Greed fell from 2.09 to 0.02, mean-reversion lost money. Only one idea kept working on data it had never seen: follow the trend. And running it on all four majors (BTC, ETH, BNB, SOL) beat any single coin — score up to 2.53. That's the strategy.
The whole rule is simple. Each coin has one line — its moving average. Price above the line? Hold the coin. Below? That part goes to stablecoins. Hold every coin that's trending, in equal parts. That's it.
Same $1, 2023 to 2026. The strategy vs just holding the same four coins. Pick the window:
$1 over time (log scale). Green = the strategy; grey = the same four coins, just held. Note: 2023–2026 was a bull market, so the returns look big — what matters is the smaller drops and the score on new data, not the headline number.
| Setting | Return · 3.4y | Per year | Worst drop | Score · new data |
|---|---|---|---|---|
| Trend majors · 50-day | +4101% | 198% | −13% | 3.23 |
| Trend majors · 100-day | +1945% | 141% | −19% | 2.53 |
| Trend majors · 200-day | +785% | 89% | −20% | 1.64 |
| Hold the 4 coins (no timing) | +254% | 45% | −61% | 0.09 |
Ask it now, with live CoinMarketCap data:
Right now all four are below their lines, in Extreme Fear. So: 100% stablecoins. It waits.
Copy the skill folder into your skills folder:
cp -r skills/trend-regime /path/to/your/skills/directory/Then connect CoinMarketCap (free key at pro.coinmarketcap.com):
{
"mcpServers": {
"cmc-mcp": {
"url": "https://mcp.coinmarketcap.com/mcp",
"headers": { "X-CMC-MCP-API-KEY": "your-api-key" }
}
}
}Then just ask: Which majors are trending right now? or /trend-regime
Not just an alert — an agent with its own wallet that makes the trades itself, no human needed. How to build it on BNB Chain:
Give it a wallet. Make a new key and add your money plus a little BNB for gas. Use only what you can risk. The private key stays in the agent's secret settings.
Run it once a day. Daily is enough — the rule uses a daily average, so checking more often changes nothing. Each run reads the four prices and their averages from CoinMarketCap, then decides the target mix in code.
Check what it holds. With viem it reads the wallet's balances — so it knows what it has versus what the rule wants.
It trades. For each coin that changed, it signs and sends a PancakeSwap trade with its own key. No alert, no human.
It's cheap. It only trades when a coin crosses its line — a few times a year. So gas is rare, and the daily check is basically free.
// daily cron — the agent's OWN key signs every swap, no human in the loop
const account = privateKeyToAccount(process.env.AGENT_PRIVATE_KEY);
const target = decide(pricesAndMAs); // the rule, in code
const held = await readBalances(account); // current on-chain holdings
for (const coin of ["BTC", "ETH", "BNB", "SOL"])
if (target[coin] !== held[coin]) // only on a flip
await pancakeSwap(account, coin, target[coin]); // signs + sends the txBuilt on Vercel Cron + Functions, the CoinMarketCap MCP, and viem. A key holding real money is a real risk — test it without funds first, use a small separate wallet, set slippage limits, and a kill switch.
The 30-strategy test, the new-data check, the skill, this site, and the agent — all built with Claude Opus 4.8. Why it's the right model for this:
It codes and uses tools. It ran the backtests itself and connected CoinMarketCap — working with real numbers, not guesses.
It thinks long. It kept the whole research in mind — 30 strategies, old data vs new data — across a long build, without getting lost.
Honest, not hype. It insisted on testing with new data, warned about overfitting, and refused to use numbers it couldn't check. (It hides failures only ~3.7% of the time, per Anthropic.)
From idea to rule. It turned a vague "follow the trend" into a clear, exact, testable rule.
Claude does the research and the building. The strategy stays a simple rule in code. More on Claude Opus 4.8 →
This is a research project built for a hackathon. It is not financial or trading advice, and not a recommendation to buy or sell anything. You alone are responsible for your decisions, your money, and your risk.
Trend-following is not risk-free. In choppy markets it gets whipsawed — it sells low and buys back high, losing a bit each time. It always misses the exact top and bottom. The backtest uses daily Binance data with costs; real slippage, gas, and timing will differ. And 2023–2026 was mostly a bull market, which makes any trend strategy look good — those returns will not repeat.
Testing on new data lowers the risk of overfitting, but doesn't remove it: we tried 30 strategies, and the more you try, the easier a lucky one slips through. We guard against this by requiring a whole group to work on new data — but no backtest can promise the future. Past results don't predict future results.
Do your own research. Only use money you can afford to lose.