Atlas backtests with AI agents
- AlgoAlpha launched Atlas on April 13, pitching an AI trading agent that searches pre-computed backtests from plain-English prompts instead of manual strategy tuning. - Atlas says it scans thousands of strategies across 85 symbols, 3 timeframes and 20,000 historical bars, then ranks results by profit, Sharpe and drawdown. - The pitch lands amid familiar warnings that backtests can flatter strategies fitted to old data, not future markets. (investor.gov)
Backtesting is a dress rehearsal for a trading strategy: you run old market data through a rule set and see how it would have performed. AlgoAlpha says its new Atlas product turns that process into a chat prompt. (customers.algoalpha.io) (docs.algoalpha.io) AlgoAlpha introduced Atlas on April 13 as an “AI backtesting agent” that lets users ask for setups like a low-drawdown Bitcoin strategy on a 15-minute chart. The company says Atlas then searches a library of pre-computed strategy results and returns ranked candidates. (customers.algoalpha.io) (algoalpha.io) The company’s pitch is speed and scale. Atlas says it covers 85 liquid symbols, three timeframes — 5-minute, 15-minute and 1-hour — and backtests built on 20,000 bars of historical data. (algoalpha.io) (docs.algoalpha.io) Instead of building rules by hand, users type a goal in plain English and Atlas filters for metrics like win rate, net profit, Sharpe ratio, max drawdown and trade count. AlgoAlpha says users can refine the same search with follow-up prompts such as “same but on 15m” or “prefer lower drawdown.” (algoalpha.io) (customers.algoalpha.io) The underlying idea is not that the model invents a strategy from scratch in real time. AlgoAlpha’s docs say Atlas searches a pre-computed database of strategies, then uses an AI interface to narrow the list and explain the results. (docs.algoalpha.io) (algoalpha.io) That matters because backtests are easy to make look cleaner than live trading. The Securities and Exchange Commission’s investor education site warns that performance claims can be presented in many ways and should be examined carefully before investors rely on them. (investor.gov) In trading, the main hazard is overfitting — building a rule set that matches quirks in old data instead of a repeatable market pattern. QuantStart, a long-running quantitative trading publication, says backtests can be distorted by optimisation bias and other statistical errors if developers keep tuning until the historical curve looks good. (quantstart.com 1) (quantstart.com 2) AlgoAlpha tries to answer that criticism in its own marketing. Its Atlas page says results are “validated across market regimes” and says the tested data includes bull, bear and ranging conditions. (algoalpha.io) The limits are visible in the product description too. Atlas currently works with strategy shapes of one trigger plus zero to two state filters, using AlgoAlpha’s SSA and ILPAC indicators across long, short or both directions. (docs.algoalpha.io) Atlas is also not a free public tool. AlgoAlpha’s documentation says access requires an active subscription and that Atlas is included in its VIP Bundle plan. (docs.algoalpha.io) So the real shift is not that artificial intelligence solved trading. It is that one vendor is packaging large libraries of historical tests behind a chatbot, making strategy search faster while leaving the old problem intact: a backtest is still only a rehearsal. (customers.algoalpha.io) (investor.gov)
