If you have ever asked an AI assistant whether your backtest is any good, you already know the problem. It tells you the Sharpe ratio looks strong. It lists some considerations. It does not ask how many configurations you tried before this one, it does not check whether your returns are just HML and momentum in a costume, and it certainly does not compute the deflated Sharpe before saying something encouraging.
We built a free Claude skill that does all three, and cites a page number when it does.
Download experienced-quant-investor.skill (free)
What’s inside
The skill was built from 71 primary papers on quant trading and factor investing: Fama-French, the AQR series on quality, low beta and momentum, Lopez de Prado on overfitting and causality, Damodaran on equity risk premiums and returns on capital, Meucci on portfolio construction, Kelly-Xiu on machine learning.
It is not a summary of that literature. Claude already knows those papers approximately from training, and approximate knowledge is the failure mode we were trying to fix. A skill that restates famous papers adds nothing and creates a surface where the restatement can be wrong without anyone noticing.
So the skill carries only what general knowledge handles badly:
anchors.md, a table of what the literature actually reported. Sharpe ratios, t-statistics, sample periods, and whether the figure is gross or net. Every number was read out of the paper’s own text and carries the page it came from.debates.md, eight real contradictions between papers in the same corpus, each with a stated position and the evidence that would reverse it.audit.md,critique.mdanddesign.md, the three working modes, each with a fixed output format and a stop rule.methods.mdandcosts.md, on when to run each test and how to read the result honestly.- Twelve runnable Python scripts covering deflated Sharpe, probability of backtest overfitting, purged cross-validation, Newey-West factor attribution, cost and capacity analysis, and a static scanner for look-ahead bias in backtest code.
The scripts depend only on numpy, pandas and scipy. Forty-one self-tests, all passing.
How to install it
The download is 88 KB and installs to 216 KB. It contains no third-party paper text.
In Claude (claude.ai and the desktop app)
- Download
experienced-quant-investor.skillwith the button above. - In Claude, open Settings, go to Capabilities, then Skills.
- Click “Upload skill” and select the file.
Note that skills require a paid plan with code execution enabled.
In Claude Code (CLI and IDE)
- Download and unzip the file. It is a standard zip archive.
- Move the folder to
~/.claude/skills/experienced-quant-investor, so thatSKILL.mdsits at the top level of that directory. On Windows the path isC:\Users\YOURNAME\.claude\skills\experienced-quant-investor. - Start a new session. The skill loads on its own when you mention a strategy, a backtest, a factor, or a Sharpe ratio.
Nothing else to configure. The scripts run from scripts/ inside that folder.
What to ask once it’s installed
You do not need to name the skill. Describe the work and it triggers.
- “I have a long-short value and momentum strategy on US small caps, 2010 to 2024, Sharpe 1.9, 40% annual turnover. Is it real?”
- “Run a proper audit on these monthly returns.” Attach a CSV with a date column and a return column.
- “Check this backtest for look-ahead bias.” Point it at a file or a folder.
- “Is my alpha actually alpha?” It runs the regression against live Ken French data and tells you what is left once the factors are paid for.
- “What Sharpe did Betting Against Beta actually report, and over what sample?”
On the first prompt, the skill will tell you it only has headline statistics, state what that ceilings the conclusion at, and ask for the one file that would fix it. That behavior is deliberate. Most backtest reviews fail because they deliver a confident verdict on evidence that cannot support one.
How we checked the work
Nothing in this skill comes from an AI’s memory.
Every paper was extracted to plain text with a page marker on every page, then read against a fixed schema that required a page citation for every number. An automated checker then verified each figure against the page it cites. All 61 numbers in the anchors table confirm. Every script ships with a self-test, and several reproduce published worked examples exactly. The minimum backtest length calculation returns 5.0 years for 45 trials and 1.9 years for 7, which are the two examples given in the source paper.
The extraction itself needed fixing before any of that was trustworthy. PyMuPDF, the common Python library for this, silently drops the minus sign in PDFs whose embedded fonts lack a Unicode mapping for it. A reported -0.49 comes out as 0.49, with no error and no visible trace. It affected 68 of the 70 text-layer papers. We rebuilt the pipeline to take layout from pdftotext and resolve each missing glyph against pypdf’s glyph names, which restored 3,602 minus signs. Anything still ambiguous is written as a visible question mark rather than a guess.
Where a paper’s results are embedded images and no number survived extraction, the skill says so and refuses to present a figure as verified. Three papers in the corpus are in that category.
Get the skill
Give Claude the actual numbers and the tests that go with them, and spend your time on research instead of on the parts of a backtest that were never going to survive.
Download experienced-quant-investor.skill (free)
The skill is MIT licensed and open source. It is an independent community project, not affiliated with or endorsed by any of the authors, journals or firms whose research it cites.
If you work in Portfolio123, our free Portfolio123 Claude Skill is the companion to this one: that skill writes screens and ranking systems the platform accepts, and this one tells you whether the result is worth trading.
This skill is for education and research. It analyzes strategies, methods and published evidence. It does not constitute investment advice, it does not recommend securities, and it does not connect to brokerage accounts.