LibraryReference articles

Quant vs data scientist, in quantitative finance

Reference article

Take a signal whose correlation with tomorrow's return is 0.05. Squared, that is 0.0025: it explains a quarter of one percent of what happens. In many data science settings a model like that is thrown away. In a market it can call the direction right about 51.6 times in 100 instead of 50, and the question becomes whether that small edge is real, and whether it survives the cost of trading. That change of question is most of the difference between the two jobs.

On this page, a data scientist is someone who uses statistics and code to learn from data and to build models that make predictions, often about customers, products or operations. A quant works in quantitative finance: work where people use maths, statistics and code to price things, manage risk and make trading decisions in markets. Of the four quant jobs, the closest to data science is the researcher, who looks for patterns in market data and tests whether they are real. See the four quant jobs for the others.

We do not publish pay figures on this page or anywhere on Thuztra Learn. Nothing here promises a job.

What the two jobs share

Both jobs use the same toolkit:

If you already do these four things well, you are not starting from zero. The differences below are about the data you point them at.

Difference 1: the signal is tiny

A correlation is a number from minus one to one that says how closely two quantities move together. Zero means no straight-line link. One means they move in lockstep.

Take a signal, meaning any number known today, whose correlation with tomorrow's return is 0.05. A common way to judge a prediction is the share of the ups and downs it explains, called R-squared. For one signal it is the correlation squared: 0.05 times 0.05 is 0.0025, a quarter of one percent. In many data science settings, a model that explains a quarter of one percent of the variation is a model you discard.

Now look at it as a call on direction. Suppose the signal and the return both follow the bell-shaped normal curve together. A standard result for that case gives the chance that the two have the same sign as one half plus the arcsine of the correlation divided by pi. With a correlation of 0.05 that is about 0.516. So the signal calls the direction right about 51.6 times in 100, against 50 for a coin toss. We computed these numbers ourselves.

Two things follow. First, small edges are the normal case in markets, so a quant reads a weak fit differently. Second, the edge is so small that proving it exists takes a lot of data. The usual noise in an estimated correlation is about one divided by the square root of the number of observations. To see a correlation of 0.05 at twice its own noise, you need the square root of the count to reach 40, which means about 1,600 daily observations. That is a little over six years of trading days, for one signal on one market.

Difference 2: the market reacts to you

A recommendation model usually does not get worse because a rival built the same model. A trading signal can. If many people trade on the same pattern, their trades move prices until the pattern is smaller or gone. In finance, an edge like that is called alpha: a return that a simple benchmark does not explain.

There is published evidence for this decay in stocks. R. David McLean and Jeffrey Pontiff studied 97 variables that earlier papers had shown to predict stock returns. In their Journal of Finance paper (2016), portfolio returns built on those variables were 26% lower out of sample and 58% lower after publication. The authors conclude that investors learn about mispricing from academic papers.

In our view this is the deepest difference between the two jobs. A data scientist can often assume the world will behave tomorrow much as it did in the training data. A quant has to assume that a good pattern is under attack.

Difference 3: the order of the rows is a fact

Market data is a time series: values recorded in time order, where the order itself carries meaning. Two problems come from that.

Leakage through time. Look-ahead bias means a test uses information that was not yet known at the moment of the decision. A common case is a company's yearly figures, stamped with the end of the year they describe but often published weeks later. A test that uses them on the stamp date has quietly seen the future. Shuffling rows for a random train and test split, which is routine in many data science tasks, causes the same leak.

Rules that drift. Markets change their behaviour: how much prices move, which assets move together, how much a trade costs. A model fitted on one stretch of history can meet a different world on the next.

Difference 4: many tries make luck look like skill

Here is the second worked example. The Sharpe ratio is a strategy's average return divided by the standard deviation of its returns, scaled to a year. It is a score often quoted for a strategy.

Suppose a strategy truly has no skill, so its true Sharpe ratio is zero. Measured on five years of daily data, the estimate still wanders. A standard approximation for a true value near zero puts its noise at about one divided by the square root of the number of years: one divided by the square root of 5, about 0.447.

Now test 20 such strategies that have nothing to do with each other, and keep the best. We simulated the best of 20 draws from a standard bell curve and got about 1.87 on average. Times 0.447, the best of 20 zero-skill strategies shows a Sharpe ratio of about 0.84 by luck alone. Test 100 and the best is about 2.51 times 0.447, about 1.12. Those are our own numbers. They assume independent strategies and bell-shaped returns, which real markets do not quite give you.

This is the problem of multiple testing: the more ideas you try, the better the best one looks, whether or not any of them works. Campbell Harvey, Yan Liu and Heqing Zhu make the case at the scale of a whole field. In their paper on hundreds of published factors, they argue that a newly found factor should clear a t-ratio of about 3.0, not the usual 2.0. A t-ratio is an estimate divided by its own noise.

Tools exist to charge for the search. One is the deflated Sharpe ratio, which marks a result down for the number of trials behind it. Our article on testing three hundred ideas at once works through the maths.

Difference 5: how a result is judged

A data science model is often judged by accuracy on held-out data. A quant result is judged by what would have happened if you had traded it.

A backtest that looks good can still be luck, and past results do not tell you the future.

Quant researcher vs data scientist: the short version

Data scientist (typical)Quant researcher
Typical targetA customer, product or operations outcomeA future return, price or risk
Usual signal strengthOften strong enough to see by eyeUsually small; the example above explains a quarter of one percent
Does the target react?Mostly notYes: others trade the same patterns
Data orderOften can be shuffledKept in time order; shuffling leaks the future
Main dangerA model that does not generaliseLuck from many tries, and costs

The table is our summary of the five differences above. Both jobs vary a great deal between employers, so read each job description yourself.

Moving from data science to quant: what to learn next

We suggest this order. Each step names a module of the Core quant curriculum or a role path on Thuztra Learn.

  1. Markets as mechanisms. What an exchange, an order and a share are. Module: Markets from zero.
  2. Desk statistics. Estimator bias, correlation traps, multiple testing, heavy tails and the Sharpe ratio. Module: Level 1: Estimates, tails and desk numbers.
  3. Inference for research. Regression with financial data, multiple testing and cross-validation when rows are in time order. Module: Statistical inference for research, on the researcher path.
  4. Data that does not leak. Point-in-time data, survivorship bias and reproducible pipelines. Module: Market data and research infrastructure, on the developer path.
  5. Machine learning, with its limits. Trees, boosting and penalised regression, and where they fail on market data. Module: Machine learning, and where it does not help.

For the route by role, read the quant roadmap by role.

Where Thuztra Learn fits for a data scientist, and where it does not

Thuztra Learn is our product. It teaches the maths of quantitative finance from zero in short lessons, and the modules named above are part of it. It does not teach product analytics or the experiment design used to test product changes, which are large parts of many data science jobs. It is not a degree, and it does not place anyone in a job. Membership is a paid subscription: see what Thuztra Learn is and what it costs.

To practise interview questions, see the bank of interview questions and the quant puzzles.

A first step from data science

Take one dataset you already know that has a date on every row. Split it two ways: once at random, and once by date, fitting on the earlier part and testing on the later part. Compare the two test scores.

You know it worked if the date split gives the lower score, or the same one. If it scores lower, the gap is a rough measure of how much the random split was leaking the future. Two things go wrong most often. The dataset has no time order that matters, such as a fixed survey: pick one with prices, sales by day or sensor readings instead. Or the date split scores higher: check that no feature was computed using rows from after its own date.

Fine print: McLean, Pontiff, Harvey, Liu and Zhu are named only as authors of published research. They have no link to Thuztra. Education only, not advice.

In short

The toolkit is shared. What changes is the data: weak signals, rules that drift, other people who trade away what you find, and many tries that make luck look like skill. In our view, a data scientist who learns to test for those four things is most of the way to quant research.

Terms used here

Related

The four quant jobs

Researcher, trader, developer and risk, described by what each actually does all day. One rule followed through all four desks, and the places where the four of them disagree.

The quant roadmap by role, in quantitative finance

What to learn, in order, for a quant trader, researcher, developer or analyst. One shared start, four branches, and a plain list of which steps Thuztra Learn teaches and which it does not.

What a backtest is, in plain words

Ten numbers, one written rule, and the whole test done by hand. Then the harder half: why a past that was kind to a rule proves very little.

The Sharpe ratio

Excess return divided by volatility, the square-root-of-time scaling, the standard error of the estimate, and the four ways the number lies about a real strategy.

Hypothesis testing when we test three hundred things

What a p-value means and what it does not, how the significance threshold has to move with the number of trials, and where the widely quoted hurdles in finance came from.

Point-in-time data and look-ahead bias

A vendor's current fundamentals table is not what was knowable then. The three dates every datum carries, the join that respects them, and why no lag rule fixes a restatement.

More on these topics

Topics

Also in the library