There is a particular request that every AI-in-finance product seems built to invite: "find me a strategy that makes money." It is a natural thing to want, and a language model will happily oblige. Ask it for an edge and it will produce one — a plausible-sounding entry rule, a confident backstory about why it works, maybe a tidy number attached. The narrative will be fluent. It will also, in all likelihood, be fiction.
Quantin is QuanterLab's built-in assistant, and it is deliberately not that machine. It is wired to help you think, explain the math you are already using, and argue against your own ideas — not to hand you a backtested edge. This article is about why that distinction matters, and how Quantin is actually built to honor it.
Feynman (1974), in his "Cargo Cult Science" address, put it plainly: the first principle is that you must not fool yourself — and you are the easiest person to fool. An eager, articulate assistant does not remove that danger. It amplifies it, because a confident answer feels like evidence. Quantin's job is to make fooling yourself harder, not easier.
Why "find me alpha" is the wrong question to ask an LLM
A language model generates text that is statistically likely given its prompt. When you ask it for a profitable strategy, the most likely continuation is a strategy that sounds profitable — because that is the shape of the millions of confident write-ups it has read. It has no access to whether the thing works out of sample, no idea how many other rules you implicitly tried before landing here, and no stake in being wrong. It will not tell you the trade count is too small to mean anything. It will tell you a story.
This is exactly the failure mode honest quantitative research is organized to defeat. Harvey and Liu (2015) showed that once you account for the sheer number of factors that have been tested, most published "discoveries" do not clear a defensible bar — the multiple-testing problem makes spurious results the default, not the exception. Bailey, Borwein, López de Prado and Zhu (2014) made the related point in their "Pseudo-Mathematics" essay: with enough trials, a backtest with a beautiful in-sample Sharpe can be manufactured for almost any series, and the mathematics will look rigorous while meaning nothing. An LLM asked for an edge is, in effect, a tireless trial-runner with a gift for narrative. That is the worst possible combination for self-deception.
So the design choice underneath Quantin is simple: it should never be the thing that produces the edge. It should be the thing that helps you scrutinize one.
The honest posture: explain, surface, refute
Popper (1963) argued that a claim earns its standing not by how much evidence we can pile up for it, but by how seriously we try to break it — a hypothesis is scientific to the degree it is falsifiable, and our job is to attempt the falsification. That is the posture Quantin is built to support. Across its modes it does four things, none of which is "give you an answer to trade":
- Explain the math. In its Mathematician mode, Quantin renders formal derivations — the Ornstein-Uhlenbeck process, a Kalman filter, GARCH conditional volatility — with LaTeX, grounded in how QuanterLab actually implements them rather than generic textbook theory. The point is comprehension, not a recommendation.
- Surface assumptions. Its base instructions tell it that when a result smells fragile — a high Sharpe on a handful of trades, a single best-cell parameter pick, no walk-forward — it should gently raise the concern rather than gloss over it.
- Propose how you could be wrong. The Skeptic mode is an adversarial reviewer by design. Its written brief is blunt: don't be a yes-man, push back, find the weakness. It is told to identify the most fragile claim you made and probe it — why this parameter and not its neighbors? your stop is inside the noise, defend it. fourteen trades is below significance, what's your plan? — and after a few rounds to deliver a verdict of Defended, Borderline, or Concerning.
- Check your reasoning against your own work. In Librarian mode it has read-only access to your saved backtests and configs, scoped to your account, so it can point out that five of your eighteen configs are near-duplicates, or that a result you are leaning on has not been refreshed in months.
Quantin presents five modes — Tutor (explain concepts, point to Knowledge Base articles), Librarian (examine your saved work), Mathematician (formal derivations), Skeptic (adversarial review), and Agent (drive the platform's tools end-to-end). They share one cached system prompt and differ only in a short behavioral extension stacked on top. The "lens" you choose changes how Quantin argues with you — not whether it will invent an edge for you. It won't, in any of them.
What Quantin will and won't do
These are not aspirations in a manifesto; they are constraints written into the assistant's actual instructions. A few worth knowing about:
It will not recommend trades or claim future returns. Every relevant mode repeats the same rule — explain methodology, never bless a specific trade, never predict a number. The Agent mode, which can actually launch backtests and scans on your behalf, closes with the line that it orchestrates research while you decide what to trust. The boundary is the whole point.
It will not invent features. Quantin's knowledge of what the platform can do comes from a single, self-maintaining capability catalog folded into its prompt, with an explicit directive: a feature listed there exists and may be explained; a feature not listed must not be claimed. A continuous-integration test fails the build if a new Primitives capability ships without being added to that catalog. This is a small thing that prevents a large class of confabulation — the assistant cannot smoothly describe a tool that does not exist, because it is told, on every request, exactly which tools do.
It will push back when you ask it to cut a corner. If you ask the Agent to deploy something fragile or skip walk-forward validation, it is instructed to refuse and explain why, the same way the Skeptic would. The platform's anti-overfit discipline is meant to be conversational, not just a setting you can click past.
It pairs its critique to the right paradigm. This is a subtle one worth dwelling on. A naive assistant recommends "run a walk-forward" for any strong-looking result. But walk-forward does not apply to a point-in-time factor backtest, a portfolio optimizer's forward-looking estimate, or a present-day screen — recommending it there is itself a kind of error. Quantin is fed per-module guidance so its scrutiny fits what you are actually looking at: probe sample-period regime coverage and factor decay on a factor backtest, probe the covariance estimator and lookback on an optimizer, and so on. Honest skepticism has to be correct skepticism.
None of this makes Quantin a source of truth. It is a competent reader of your screen and a willing opponent of your conclusions — and it can still be wrong, miss things, or state something with more confidence than it has earned. Treat it the way you would treat a sharp colleague who has not seen your data before: useful for finding holes, never the final word. The validation lives in your walk-forward, your out-of-sample reveal, your forward test — not in a chat reply.
The useful inversion
There is a temptation, when you have an idea you like, to go looking for confirmation — and an assistant that gives it to you feels wonderful right up until the live results disagree. Aronson (2006) spent a whole book on the discipline of subjecting technical-analysis claims to genuine statistical testing precisely because the field had so often skipped it; the recurring lesson is that what survives honest scrutiny is a small fraction of what feels compelling. The most valuable thing an AI in this seat can do is run the other way: take the idea you are attached to and try to take it apart.
That is the inversion Quantin is built around. Used as an oracle, it would be a fast, fluent way to fool yourself. Used as a reviewer — explain this to me, what am I assuming, how could this be wrong, where is the weakest claim — it does the one thing the easiest-person-to-fool problem demands. The tool is honest only if you ask it honest questions. It is set up to make those the easy ones.
- An LLM asked to "find alpha" will confabulate a fluent, confident narrative — the precise failure mode honest research exists to prevent.
- Quantin is built to explain the math, surface fragile assumptions, and argue against your idea (Skeptic mode is an adversarial reviewer by design) — not to hand you a backtested edge.
- Its instructions hard-code the limits: never recommend trades, never claim future returns, never invent a platform feature (a CI test enforces the capability list), and push back when asked to skip validation.
- Its critique is paradigm-aware: it recommends walk-forward only where walk-forward applies, and probes the right weaknesses for factor backtests, optimizers, and screens.
- The honest use is the inverted one: ask Quantin to refute your idea, not to bless it. The validation lives in your own out-of-sample work.
Quantin is, in the end, a small instrument pointed at a hard human problem. It cannot make your strategy true, and it will not pretend to. What it can do is sit across the table and ask the questions you would rather not — which, if you are serious about not fooling yourself, is exactly the help worth having.
Further Reading
Foundational papers
- Feynman, R. P. (1974). Cargo Cult Science (Caltech commencement address). Engineering and Science, 37(7), 10–13.
- Popper, K. R. (1963). Conjectures and Refutations: The Growth of Scientific Knowledge. Routledge & Kegan Paul.
- Harvey, C. R. & Liu, Y. (2015). Backtesting. Journal of Portfolio Management, 42(1), 13–28.
- Bailey, D. H., Borwein, J. M., López de Prado, M. & Zhu, Q. J. (2014). Pseudo-Mathematics and Financial Charlatanism: The Effects of Backtest Overfitting on Out-of-Sample Performance. Notices of the AMS, 61(5), 458–471.
Textbook references
- Aronson, D. R. (2006). Evidence-Based Technical Analysis: Applying the Scientific Method and Statistical Inference to Trading Signals. Wiley.
Related QuanterLab articles
Try it in QuanterLab
Try it in QuanterLab. Open Quantin from any module and switch it to Skeptic mode. Paste in a result you are proud of — or describe a strategy idea you are excited about — and ask it to find the weakest claim. Then try the same idea in Mathematician mode and ask it to derive the formula you are relying on. Notice that at no point does it tell you what to trade; it only helps you see your own reasoning more clearly.