Quantitative trading means making trading decisions with mathematical models built from data, not with discretionary judgement. A quant strategy turns an idea into explicit rules, such as "buy assets whose past 12-month return is positive". It tests those rules on historical data. If they hold up, a computer runs them. Most of the work, and most of the failures, happen in the testing step: a model that looks strong on past data is easy to build and much harder to make work on new data.
Quantitative, algorithmic and high-frequency: the difference
People often use these terms as if they mean the same thing. They don't.
- Quantitative trading is about how the decision is made: a model built from data decides what to trade and when.
- Algorithmic trading is about how orders are sent. EU law (MiFID II, Directive 2014/65/EU, Article 4(1)(39)) defines it as trading where "a computer algorithm automatically determines individual parameters of orders such as whether to initiate the order, the timing, price or quantity of the order" with "limited or no human intervention". Systems that only route orders are excluded.
- High-frequency trading is one kind of algorithmic trading. Under Article 4(1)(40), it has three features: infrastructure built to minimise latency (for example, servers placed at or near the trading venue), order initiation and execution decided by the system without human intervention, and high intraday message rates (orders, quotes or cancellations).
A quant strategy can rebalance once a month and send its orders by hand. A high-frequency firm almost always uses quantitative models.
How a quant strategy is built
1. A testable hypothesis
Each strategy begins with a claim about how prices behave, stated precisely enough to measure. Trend following is one of the best documented. In "Time Series Momentum" (Journal of Financial Economics, 2012), Tobias Moskowitz, Yao Hua Ooi and Lasse Pedersen found "strong positive predictability from a security's own past returns" across 58 futures and forward contracts on equity indices, currencies, commodities and sovereign bonds, using more than 25 years of data. They found the effect lasts about a year and then partly reverses over longer horizons.
2. Data
The inputs are usually price and volume. Some strategies add fundamentals, order-book data or alternative data. The data has to cover enough history and market regimes for the test to mean anything. It also has to be clean: survivorship bias, look-ahead errors and bad prints all make a backtest look better than it should.
3. Backtest
The rules are run on historical data, after costs: fees, spreads, slippage and, for leveraged products, financing.
4. Out-of-sample validation
The strategy is then tested on data that played no part in designing it. This is the step that separates an actual effect from a curve fitted to noise, as the overfitting research below shows.
5. Execution and monitoring
The model goes live, usually starting small. Live results are tracked against what the backtest predicted, and there must be a way to switch it off.
Should you trust this backtest?
Answer six questions from the article's checklist to see where a quant result is most likely to break.
Was the hypothesis stated before testing, with a reason it should work?
- Yes → next question
- No, or not sure → No testable idea yet
Do you know how many variants were tried, and whether the reported one is the best of many?
- Yes, every variant is recorded → next question
- No record of the attempts → High risk of overfitting
Does the backtest include fees, spreads, slippage and, for leveraged products, financing?
- Yes → next question
- No, or only some of them → Returns are probably overstated
Was the strategy tested on data that played no part in designing it?
- Yes → next question
- No → High risk of overfitting
Does it still work with slightly different parameters?
- Yes → next question
- No, results fall apart → High risk of overfitting
Is there a kill switch, and is live performance tracked against the backtest?
Without a precise claim about how prices behave, stated before testing, a good backtest may be a pattern found in noise. State the hypothesis first, then test it.
The research the article cites finds that strong simulated performance is easy to reach after trying a small number of configurations, and the more configurations tried, the more likely the backtest is overfit. Out of sample, an overfit strategy can lose money, not just earn nothing.
A backtest that leaves out fees, spreads, slippage or financing makes the strategy look better than it should. In leveraged crypto products, financing costs and liquidation risk matter even more.
A correct model can still fail through bad code or weak controls, as the 2012 Knight Capital incident showed. Supervisory guidance stresses the ability to disable an algorithm quickly and to monitor live trading.
The result clears every check in the article's list. That lowers the chance of a common failure, but it does not prove the effect will last on new data.
This walks through the article's checklist only: it does not test any strategy, measure how overfit a result is, or say whether to trade it.
Where quant strategies fail
Backtest overfitting
The best-known warning is "Pseudo-Mathematics and Financial Charlatanism" by David H. Bailey, Jonathan Borwein, Marcos López de Prado and Qiji Jim Zhu (Notices of the AMS, May 2014). Its findings:
- "High simulated performance is easily achievable" after "backtesting a relatively small number of alternative strategy configurations."
- "The higher the number of configurations tried, the greater is the probability that the backtest is overfit."
- Analysts rarely report how many configurations they tried, so investors cannot judge how overfit a result is.
- Under memory effects, overfitting can produce negative expected returns out of sample, not just zero. The authors suggest this "may be one of several reasons why so many quantitative funds appear to fail."
The practical point: keep a record of every variant you test. A strategy chosen as the best of 200 attempts needs far stronger evidence than one designed in advance.
Data mining across the whole field
The same problem shows up across academic research. In "…and the Cross-Section of Expected Returns" (Review of Financial Studies, 2016), Campbell Harvey, Yan Liu and Heqing Zhu argue that the usual threshold for a new return factor (a t-ratio above 2.0) is too low once you account for how many factors researchers have tested. They propose a threshold of about 3.0, and conclude that many published findings are likely false.
Operational risk
A correct model can still fail through bad code or weak controls. On 1 August 2012, Knight Capital's router sent more than 4 million orders in the first 45 minutes of trading while trying to fill 212 customer orders. It traded more than 397 million shares, and the firm lost more than $460 million. In 2013, the SEC charged Knight with violating the market access rule (Rule 15c3-5), and the firm agreed to pay a $12 million penalty.
What supervisors expect
FINRA Regulatory Notice 15-09 (26 March 2015) sets out supervision and control practices for firms that trade algorithmically, grouped into five areas:
- general risk assessment and response
- software and code development and implementation, including the ability to disable an algorithm quickly
- software testing and system validation, done independently of development
- monitoring of live trading systems
- compliance, including monitoring how algorithms interact with each other
The notice was written for broker-dealers, but the same controls make sense for anyone running automated strategies.
Market-wide effects
SEC staff examined this question in the Staff Report on Algorithmic Trading in U.S. Capital Markets (August 2020), which Section 502 of the Economic Growth, Regulatory Relief, and Consumer Protection Act of 2018 required. The report reviews how algorithmic trading affects liquidity both in normal markets and under stress.
Quantitative trading in crypto
Crypto markets trade around the clock, so strategies need automated monitoring and risk limits that work when nobody is watching. The research suggests that some quant ideas do carry over. In "Risks and Returns of Cryptocurrency" (Review of Financial Studies, 2021), Yukun Liu and Aleh Tsyvinski report:
- a strong time-series momentum effect in cryptocurrency returns
- proxies for investor attention strongly forecast future returns
- returns are exposed to network factors (user adoption) but not to production-cost factors
The usual cautions apply even more strongly here. Histories are short, and the data comes from many venues of uneven quality. Leveraged products add financing costs and liquidation risk, and a backtest that ignores them will overstate returns. A short history also means fewer independent market regimes to test against, which makes overfitting easier.
A short checklist before trusting any quant result
- Is the hypothesis stated before testing, with a reason it should work?
- How many variants were tried, and is the reported one the best of many?
- Are costs, slippage and financing included?
- Was it tested on data held out from design?
- Does it still work with slightly different parameters?
- Is there a kill switch, and is live performance tracked against the backtest?
