What quantitative research actually is

A student loads five years of daily prices for a single stock, plots a 20-day and 50-day moving average, and back-tests the rule “buy when the fast average crosses above the slow one.” The rule returns 18% a year over the sample, comfortably ahead of buy-and-hold. The student calls this quantitative research. It is closer to a coin that has landed heads five times in a row: an observation, not yet a finding.

The distinction matters because the word “quantitative” describes the tools — arithmetic, statistics, code — not the standard of proof. A spreadsheet formula and a peer-reviewed factor model are both quantitative. What separates them is whether the number in front of you has survived anything designed to break it.

A definition that does actual work

Quantitative research in finance is the disciplined process of turning a question about markets into a claim that can be checked, using data, an explicit method, and a rule for deciding whether the result would also have appeared by chance. Four things have to be present, or the word does not apply:

1. A question stated before looking at the answer. “Does momentum predict next-month returns in large-cap equities?” is a question. “Look at this chart — momentum works” is a caption.
2. Data with a defined boundary. Which securities, which dates, which source, which adjustments. Change any of those and the study is a different study.
3. A method that does not know the answer in advance. The same procedure, run on a different slice of the same kind of data, should be expected to produce a comparable result if the effect is real.
4. A criterion for failure. Before the test runs, there has to be a stated way the result could come back negative — otherwise every outcome, including noise, gets reported as success.

The moving-average example above fails on the fourth point. Nobody decided in advance what result would have counted as “this doesn’t work.” The rule was tried, it produced a big number, and the big number became the finding. That ordering — try, then decide it counts — is the single most common way research becomes indistinguishable from luck.

The shape of the process, not just the tools

Quantitative research in an equity or multi-asset context tends to move through the same broad shape regardless of what is being studied: define the universe of securities under study, look at what the data actually contains before touching a model, prepare that data so a later step cannot see information it should not have, build the features a model will be asked to learn from, fit and compare candidate models, turn a model’s output into a rule that can be tested as a strategy, and finally assemble everything into evidence someone else can check. A later article walks this loop stage by stageStationarity A property of a time series whose statistical characteristics — mean, variance, autocorrelation — do not change over the sample. — the point here is that each stage constrains the ones after it. A universe drawn with hindsight (only the stocks that survived to today) inflates every statistic computed from it before any model is even chosen. A look-aheadLook-ahead bias Using information in a calculation that would not actually have been available at the time the decision it informs was made. error introduced during data preparation cannot be fixed by a better model later; it has already been baked into the labels the model is trying to learn.

This is why the research loop, end to end treats research as a pipeline with dependencies, not a checklist of independent tasks. Skipping a stage does not remove its risk; it just moves the risk downstream, where it is harder to see and easier to mistake for skill.

Evidence versus a good story

A number becomes evidence only once it has been exposed to the possibility of being wrong. Two researchers can report the same headline and mean very different things by it.

A Sharpe ratioSharpe ratio Annualised excess return over annualised volatility, the standard reward-per-unit-of-risk statistic for a strategy or portfolio. measures excess return per unit of risk taken:

$\text{SR} = \frac{\mu - r_f}{\sigma}$

where $\mu$ is annualised return, $r_f$ is the risk-free rate, and $\sigma$ is annualised volatility. Researcher A tries one moving-average crossover on one stock, gets a Sharpe ratio of 1.4, and reports it as “an 18%-a-year strategy.”

Researcher B tries 60 variations of the same idea — different lookback windows, different assets, different entry rules — keeps the best one, and reports the same 1.4 Sharpe ratio without mentioning the other 59.

Both numbers look identical on a slide. They are not the same finding. Researcher B has run what statisticians call a multiple-comparisons problem: test enough variations of anything against noisy data, and one of them will look good by chance alone. The correction for this — the deflated Sharpe ratioDeflated Sharpe ratio The Sharpe ratio adjusted for the number of configurations tried and for non-normal returns, so a result picked from many looks less impressive. — discounts the reported Sharpe ratio by the number of configurations that were actually tried, because a result chosen as the best of many is a weaker claim than a result that was the only one tried.

As a rough illustration: if 60 independent, genuinely random strategies are tested against a market with no real edge, ordinary sampling variation means the best of the 60 will very likely show a Sharpe ratio well above zero — not because it found anything, but because with 60 draws from noise, the maximum drifts upward. A single honestly-reported Sharpe ratio of 1.4 and “the best of 60 tried” reporting the same number are different pieces of evidence, and only a trial count lets anyone tell them apart. This is covered in depth in signal versus noise.

What a result has to beat, before it counts

A model that predicts next-day direction with 54% accuracy sounds better than a coin flip. Whether it is a finding depends entirely on what it is being compared against. If the market itself goes up 54% of days over the sample — which happens over long enough windows in equities — then a model that always says “up” also scores 54%, without learning anything. This comparison point is the naive baselineNaive baseline The simplest rule that requires no model at all — predict no change, predict the trend continues, predict the historical average — that any predictive claim must beat.: the simplest rule that requires no model at all (predict “no change,” predict “the trend continues,” predict the historical average). A model earns the description “predictive” only once it beats that baseline by a margin that survives out-of-sample testing, not merely on the data it was fit to.

The same logic applies to portfolio work. A portfolio that returns 9% a year is not automatically good; it has to be compared against a stated benchmarkBenchmark A stated comparison point — an index, an equal-weight allocation, a risk-free rate — that a result must be measured against before it can be called good. — an index, a simple equal-weight allocation, a risk-free rate — because 9% beats one comparison and loses to another. Quantitative research always names its comparison point before celebrating a result.

Worked example: a claim, checked

Suppose a report states: “A random forest classifier achieved 61% accuracy predicting next-week direction for a 50-stock universe, evaluated with walk-forward validation across five folds, mean ROC AUC 0.58, worst fold 0.51, against a naive baseline of 52% (the empirical up-week frequency in the sample).”

Compare this with: “Our model is 61% accurate.”

The first version is checkable. It names the comparison (naive baseline at 52%), the validation design (walk-forward, five folds), the honest number to worry about (worst fold ROC AUC of 0.51 — barely above a coin flip’s 0.50), and the metric being averaged. Someone could rerun this with the same data and method and expect a similar spread of fold results. The second version is a headline with the checkable part removed. Both describe the same experiment; only one of them is research.

Why the standard matters more in finance than elsewhere

Financial data is noisy, non-stationary, and available in limited quantity relative to how many questions people want to ask of it. A physicist testing a hypothesis can often generate more data. A quant researcher testing a hypothesis about, say, monthly equity returns since 1990 has roughly 400 monthly observations, most of them correlated with each other through shared market conditions. That scarcity is exactly why the four requirements above — a stated question, bounded data, a method blind to the answer, and a pre-declared failure criterion — carry more weight in this field than in one where more data is one experiment away.