Evidence

Give your AI agent the results of 542,100 backtests (MCP and API)

By Justin JaegerPublished Updated
Your best backtest might be luck: 199 tries, best Sharpe 1.46, luck bar 1.59
Your AI knows the trading papers, not what held up: off by 17–20 points

Your agent asks BacktestAtlas before it backtests. It skips ideas that already failed, starts where the evidence points and can tell a lucky backtest from a real edge. You get to a strategy with fewer backtests, less time and fewer tokens.

Copy your personal URL and add it to any MCP client as a remote server. The free plan gives 20 checks a day without a sign-up; setup steps per app are on the start page.

https://mcp.backtestatlas.com/u/your-token/mcp

HTTP API

Every plan can call the same tools over plain HTTP with JSON, with the same limits. List them with GET /u/<token>/v1/tools, call one with POST /u/<token>/v1/<tool>:

curl -X POST https://mcp.backtestatlas.com/u/<token>/v1/check_idea -H "content-type: application/json" -d '{"idea":"RSI mean reversion on 15-minute EURUSD"}'

Tools

  • check_idea: before a backtest, the closest tested variants and the verdict for a plain-language idea
  • get_base_rates: how often ideas of a kind held up, by mechanism, asset class, frequency or data source
  • list_dead_ends: idea classes that failed repeatedly after costs, so your agent skips them
  • get_regime_evidence: which technical rule families do better or worse in uptrends, downtrends and volatility regimes, checked on a sealed holdout
  • get_strategy_library: the distribution of 542,100 labelled in-sample backtests for any rule, symbol and timeframe
  • check_backtest: a reality check for a backtest you already have: best Sharpe expected by luck, probability the edge is real, expected live range
  • get_finding: one entry with evidence, caveats and ID

Every answer carries sample size, test window and as-of date, so your agent can cite it.

Guides