You have more hunches than you ever test. "Stocks bounce after they get oversold." "Golden crosses work." "Buy the earnings beat." Testing one of them properly means writing pandas, paying for a quant platform, or giving up, so most ideas stay opinions and never meet the data. Here is the faster way: describe the strategy, get the numbers.
Does buying the dip actually work?
Take one of the most repeated beliefs in retail investing: that stocks bounce after they get oversold. Ask Claude to test it across the whole market and hold the result up against simply being invested. Figures are from June 2026.
"Backtest buying US stocks when they're oversold with an RSI under 30. What's the 21-day forward return versus the rest of the market since 2010?"
Buying oversold (RSI under 30) vs the rest of the market
21-day forward return, 2010 to 2026:
Group Signals Win rate Median Avg*
──────────────────── ─────────── ──────── ─────── ──────
Oversold (RSI < 30) 846,672 51.9% +0.34% +0.59%
Rest of the market 18,850,011 52.7% +0.31% +0.60%
* Average is capped at the 1st and 99th percentile; the raw average is
distorted by a few extreme movers. Win rate and median are the
reliable read, and on both the oversold group did not beat the market.There is the answer, on 846,672 oversold signals across 16 years: buying the dip on RSI alone did not beat simply being in the market. A slightly lower win rate (51.9% against 52.7%) and the same forward return. That cost one question and a few seconds, not a $200 platform or an afternoon of Python, and it just talked you out of trading a folk belief. Some signals do hold up, and the value of AI backtesting is finding out which, in seconds, before you risk a dollar on one.
How it works
Three steps, and only the first is setup.
- Paste the connector URL into Claude once. It takes about two minutes and needs no API key.
- Describe the strategy in plain English: the signal, the holding period, the universe.
- Claude runs it across the history and hands back the win rate, the forward returns, and the caveats that matter.
More strategies you can test
The same approach works for any signal you can describe: a price condition, a fundamental threshold, or a mix of both. A few to paste in as they are:
"Backtest a golden cross: buy when the 50-day average crosses above the 200-day. How do the 3-month forward returns compare to the market?"
"What happened to stocks in the 20 days after they beat earnings estimates by more than 10%, over the last 15 years?"
"Do high Piotroski F-score stocks (8 or 9) outperform the market over the next year?"
For these worked out step by step, with the SQL Claude writes, see the guide to backtesting stocks with AI.
Read every backtest with these in mind:
- Survivorship bias. Delisted and bankrupt companies are dropped from the data, so backtests skew optimistic; the losers that left the market are missing.
- Outliers. A handful of extreme movers blow up the raw average, which is why the figures above cap it at the 1st and 99th percentile. Median and win rate are the steadier read.
- No trading costs. Commissions, spread, and slippage are not modeled, so real returns run lower, especially for high-turnover signals.
- Overfitting. Test enough variations and one will look good by chance. A backtest is a first check, not proof; validate out of sample before risking money.
- End-of-day US data from Tiingo, and not financial advice. NYSE and NASDAQ common stock only.
Questions
Can I backtest a stock strategy without writing code?
Yes. Describe the strategy in plain English and Claude runs it against 64 years of US market data, returning the win rate, the forward returns, and the methodological caveats. No Python, no spreadsheets, and no separate backtesting platform.
Is the backtesting free?
Shibui Finance is free. You connect it to Claude once (the free Claude plan works) and ask. There is no $200-a-month platform fee and no API key; you only need a Claude account.
How far back does the data go?
Daily prices run from 1962, with 56 technical indicators and quarterly and annual fundamentals, across roughly 9,900 NYSE and NASDAQ companies. That is enough history to test a signal across many market regimes.
How accurate are these backtests?
Treat a backtest as a first check, not proof. The data is end-of-day and survivor-biased (delisted and bankrupt companies are dropped, which flatters past results), and no transaction costs or slippage are modeled, so real-world returns would be lower. Validate any promising result out of sample before risking money.
What kinds of strategies can I test?
Technical signals such as RSI, MACD, and moving-average crossovers; fundamental factors such as the Piotroski F-score, margins, or revenue growth; reactions to earnings beats and misses; and combinations of these, across the whole US market.
Is this different from the backtesting guide?
The guide walks through how to phrase backtests step by step. This page is the short version: connect Shibui to Claude and ask. Both use the same free database.
Backtesting is one way people use Shibui. Browse 20 real-world examples, screen the market with the AI stock screener, or read the full feature guide.
Shibui Finance is free, and built by Christian Richter in Berlin.
Contact
Backtesting as part of your work and want to compare notes? chris@shibui.finance
Test the next hunch you have.
Shibui is free. Connect it to Claude on any plan and backtest a strategy in plain English, instead of coding it or paying for a platform.
Connect to Claude →