WhitmanTrading

Quantitative Analysis: Measure, Then Decide

Quantitative analysis is deciding from measurements you have taken rather than from an argument that sounds convincing. The order matters more than the mathematics - ask a question with a checkable answer, decide how you would test it, then get the data right and run the test once.

How it works

A candlestick chart of the site's shared price history. The headline on the chart reads: Deciding from measurements rather than from opinions.
Deciding from measurements rather than from opinions. Illustrative chart - not real market data.

Quantitative analysis means deciding from measurements rather than from opinions. A claim becomes a number somebody else could check, and the number decides. It is a discipline about evidence, not a category of strategy — technical analysis can be done this way or by eye.

A gently rising stretch of the long price series with an account equity curve beneath it. The headline on the chart reads: It starts with a question that has a checkable answer.
It starts with a question that has a checkable answer. Illustrative chart - not real market data.

It starts with a question that has a checkable answer.Do indicators work” is not one. “On this series, does a close above the twenty-bar high lead to a higher close ten bars later” is.

A calmly advancing stretch of the long price series with a slowly rising equity curve beneath it. The headline on the chart reads: Most of the work is cleaning data, not modelling it.
Most of the work is cleaning data, not modelling it. Illustrative chart - not real market data.

Most of the work is cleaning data, not modelling it. Missing bars, adjusted against unadjusted prices, mismatched time zones and dead instruments each produce results that look real and are artefacts of the file.

A flat, quiet stretch of the long price series with a gradually rising equity curve beneath it. The headline on the chart reads: And a simple model on good data beats the reverse.
And a simple model on good data beats the reverse. Illustrative chart - not real market data.

And a simple model on good data beats the reverse. A complicated model on dirty data can always be made to look excellent, and when it fails nobody can say which part did.

How you fool yourself

A strongly rising stretch of the long price series with an account curve breaching its limit. The headline on the chart reads: Search long enough and chance will hand you a result.
Search long enough and chance will hand you a result. Illustrative chart - not real market data.

Search long enough and chance will hand you a result. Run forty tests at a five per cent threshold and two will look significant with no effect present at all. That is arithmetic, not bad luck.

A choppy, directionless stretch of the long price series. The headline on the chart reads: So write down the test before you run it.
So write down the test before you run it. Illustrative chart - not real market data.

So write down the test before you run it. The question, the threshold, and how many variants you intend to try — because a threshold chosen after seeing the result is not a threshold. That is the line between measurement and overfitting.

A declining stretch of the long price series. The headline on the chart reads: And keep data back that the model never touched.
And keep data back that the model never touched. Illustrative chart - not real market data.

And keep data back that the model never touched. Look at the held-out block once. Adjusting between looks turns it into training data, which is why forward testing beats any split of the past.

In practice

A 72-bar candlestick section of the shared price history with an account curve shown with and without fees. The headline on the chart reads: A model without costs in it is a description, not a plan.
A model without costs in it is a description, not a plan. Illustrative chart - not real market data.

A model without costs in it is a description, not a plan. On this site’s shared history a round trip costs 2% of a median bar’s range, so a short-horizon result turns as much on the cost assumption as on the logic.

A candlestick chart with a volume histogram beneath it, with the volume histogram emphasised. The headline on the chart reads: Participation is a measurable input and mostly unused.
Participation is a measurable input and mostly unused. Illustrative chart - not real market data.

Participation is a measurable input and mostly unused. Volume sits on nearly every series and price-only models dominate anyway, partly by convention and partly because the data is less consistent across venues.

A long-horizon candlestick view of the same price series. The headline on the chart reads: A longer series is worth more than a cleverer method.
A longer series is worth more than a cleverer method. Illustrative chart - not real market data.

A longer series is worth more than a cleverer method. Sample size sets what a measurement can resolve at all, so a slow strategy needs decades before its record means anything.

A candlestick series containing several opening gaps, with the largest opening gap marked. The headline on the chart reads: The rare event is under-represented in every sample.
The rare event is under-represented in every sample. Illustrative chart - not real market data.

The rare event is under-represented in every sample. An opening gap is the clear case: what matters most is what the history holds least of, so the model is weakest where the damage is largest.

A declining stretch of the long price series, with the entry price and the level at which a stop would trigger drawn as horizontal lines. The headline on the chart reads: And the model says nothing about where the risk sits.
And the model says nothing about where the risk sits. Illustrative chart - not real market data.

And the model says nothing about where the risk sits. It measures an average across many trades; the stop loss on the trade in front of you is a separate decision, and position sizing is not a modelling question.

A candlestick chart of the site's shared price history, annotated with the round-trip cost. The headline on the chart reads: Every round trip costs 2% of a bar.
Every round trip costs 2% of a bar. Illustrative chart - not real market data.

Every round trip costs 2% of a bar. Which makes trade frequency a design choice rather than an outcome, and one to measure before the method is built.

A procedure you can actually run

Write the question first, in one sentence, naming the instrument and the horizon. “Does a close above the twenty-bar high lead to a higher close ten bars later” is testable. “Do breakouts work” is not.

Then state the test and the threshold before touching the data. Which measurement decides, what result would count, and how many variants you are allowed. Count the alternatives honestly — every parameter swept and every instrument checked is an attempt.

Then run it once, with costs included, on a block you have not looked at. Report the number of attempts beside the result. A finding from one test and the same finding from the fortieth are different objects, and only the count tells them apart.

And backtesting is the same procedure applied to a whole set of rules rather than a single question. Only the number of moving parts changes.

What quantitative analysis is not

It is not a category of strategy. It is a way of checking one.

It is not mathematics for its own sake. The arithmetic is usually simple.

It is not systematic trading. That is rules, this is evidence.

And it is not protection from being wrong. It only makes wrong findable.

When it fails

A sideways, range-bound candlestick series. The headline on the chart reads: In a flat period every measurement loses its power.
In a flat period every measurement loses its power. Illustrative chart - not real market data.

In a flat period every measurement loses its power. Inside a trading range the difference you are trying to detect shrinks below the noise, and a method that measured something real in a trend measures nothing here.

The second failure is comparing against the wrong alternative. A result measured against zero rather than the base rate flatters itself, because the drift of the market is doing the work.

A third is the summary statistic that hides the shape. A mean is fine, standard deviation less so on a fat-tailed distribution, and a Sharpe ratio says nothing about the worst day.

A fourth is treating correlation as a mechanism. Two series moving together over a chosen window is a measurement, not an explanation, and the window is usually doing the work.

A fifth is the sample that quietly survived. A history holding only the instruments still listed is the least visible bias in this work, because the failures were deleted before you measured.

And a sixth is re-testing the data you held back. Once it has informed a change it is training data, and the honest response is to say so and find more.

The original data

On this site’s shared 576-bar history the base rate for a higher close ten bars later is 54% across 566 observations, and 52% over one bar across 571 observations. Both figures are in research/series-measurements.json, produced by site/measure_series.py.

A strongly rising stretch of the long price series, cut short at the decision bar. The headline on the chart reads: Significant at five per cent, after forty tests. Real?
Significant at five per cent, after forty tests. Real? Illustrative chart - not real market data.

That 54% is the whole method in miniature. A signal advertised as right 55% of the time is a claim about one percentage point over doing nothing at all — and the round trip costs 2% of a median bar’s range, more than the edge being claimed. Measure the alternative before believing the claim.

The corpus figures say the same from the other side. Twelve videos carry “quantitative” in the title at a median of 204 views, against 211 on “win rate” at a median of 11,527, per research/broker-coverage.json. Measure the base rate for your instrument and horizon before believing any claim, then subtract the costs.

Systematic trading is what happens once the measurements become rules precise enough to run without you. Overfitting is the way this work goes wrong, and the reason the test gets written down first. And probability is the language the results are stated in, including why five per cent means less than most people assume.

What I actually do

I used to accept a setting because somebody I respected said it worked. The first time I properly measured one on my own data, the result was so much weaker than the claim that I stopped trusting anything I had not checked myself. What changed my trading was not a better method, it was the habit of asking what happens without the signal. Most claimed edges do not survive that question.

— Michael Whitman

This page is educational, not financial advice. Test every idea on your own charts before risking money.