Quantitative analysis: how a market hypothesis becomes a rule and where it breaks

Quantitative analysis is a way of making trading decisions by a rule that is formulated in advance, tested on history and described precisely enough to be executed without any judgement in the moment. Its product is not a price forecast for a particular security but an assessment of how robust an observed market property is: whether it persists across different stretches of history, across different instruments, and after costs are deducted. If the property fails at least one of these tests, there is no rule — there is a coincidence that has already happened.

What is measured here and what is not

The quantitative approach is not opposed to either the fundamental or the technical view: it takes their features and turns them into measurable quantities. In fundamental logic, a reported figure such as 3,92 is an occasion to discuss fair value; in quantitative logic it is simply a coordinate by which a security can be compared with the rest and ranked. A chart pattern in technical analysis is described in words; here it must be written as a condition that a computer will check in the same way on every bar.

Hence the main limitation of the discipline: it works only with what has been digitised. A corporate conflict, a change in the regulatory framework or the quality of the management team does not enter the model until someone devises a measurable proxy for it — and a proxy is always poorer than the phenomenon itself. This is not a defect of the method but the boundary of its applicability, and it is worth keeping in mind alongside the boundaries of the neighbouring approaches: see fundamental analysis and what technical analysis cannot do.

Where a hypothesis comes from

The order matters, and it is the opposite of the intuitive one. Not "I'll dig through the data and see what turns up", but "there is a reason why the market might pay for this property — I'll check whether it does".

Step one is to name the mechanism. Why should the edge exist at all? The answer usually comes down to one of several types: someone takes on a risk that others find inconvenient to hold; someone provides liquidity where it is in short supply; information spreads unevenly among participants; institutional constraints force large players to trade not when it suits them but when their mandate requires it.

Step two is to express the mechanism as a feature that can be calculated from the data available. Financial statements are taken from issuer disclosures, price history from equity data, and calendar effects from the events calendar.

Step three is to write down in advance which result will show the hypothesis to be wrong. A rejection criterion fixed before the test is the only thing that separates testing from curve fitting.

The backtest: what it does not prove

The errors that spoil a backtest are almost always the same. Overfitting: there are many parameters and little data, so the rule memorises noise. Look-ahead bias: the calculation includes information you could not have had at the time of the trade — financial statements before their publication date, a price adjusted retroactively, the final composition of an index. Survivorship bias: the test runs on a list of securities that have made it to today, while the delistings and the wipe-outs have dropped out of it. Period selection error: the sample starts and ends where the picture is more convenient.

The antidote is to split the data into parts, one of which takes no part in parameter selection at all, and to check how stable the result is when the settings are shifted. If a small change in a threshold destroys the effect, there was no effect. This is exactly the logic described by sensitivity analysis, and here it is not an add-on but a mandatory part of the procedure.

Costs and capacity

An edge is measured after costs, not before. Costs include commission, the spread, slippage on execution and the price impact of the order itself, if its size is comparable to the liquidity of the security. High-frequency rules are the most sensitive to this: an effect visible on daily closes is often eaten up entirely by the fact that nobody trades at those closes.

Capacity is a separate question. A rule that works beautifully on a small account in second- and third-tier stocks stops working as the size grows, because it becomes part of the price itself. Compare its behaviour in the liquid segment and in the OFZ market, where the microstructure and the mix of participants are arranged differently — the same hypothesis lives a different life there.

Risk management is part of the model, not an addition to it

A quantitative strategy sets not only the moment of entry but also the position size, the rule for cutting exposure in a drawdown, and the limit on concentration in a single issuer and sector. The sizing decision affects the final result more than the accuracy of the signal does: a rule with a moderate edge and disciplined sizing survives periods in which an aggressively leveraged rule with the same signal is wiped out.

Since a backtest describes the past, the future has to be described separately — through scenario analysis: what happens to the rule if volatility rises sharply, if part of the market closes, if the tax or infrastructure framework changes. The answers here are qualitative, but they need to be recorded before the scenario arrives.

When the quantitative approach is not needed

It pays off poorly if your decisions are infrequent and you have few features: the model is then built on a sample where there are not enough statistics for anything. It is harmful if it turns into endless optimisation without any trades — a state described as analysis paralysis. And it is useless if its conclusions are not executed: a rule that is overridden by hand every time it becomes uncomfortable is not a rule.

A candid caveat: the site does not publish specific performance estimates for individual strategies on the Russian market, and there are none in this article — nor is there data on which such estimates could be properly built. The testing is left to you, on your own data. Definitions of the terms are collected in the glossary, and the current market picture is in the news feed and the dividend calendar.