Quantitative trading turns a market idea into a number you can test before any money moves. Here is the research method, step by step, with a published crypto momentum signal fitted on almost two years of Hyperliquid data and tested on the sixteen months after, and what the test showed.
November 16, 2026
19
min to read
The short answer
Quantitative trading means making trading decisions from statistical evidence instead of judgment. A quant states a hypothesis, turns it into a signal computed from data, tests whether the signal predicts returns, and charges every cost. Then comes the step that matters most: running it once on data it was never fitted on.
Quantitative trading is trading on evidence you can measure. Instead of “this chart looks strong,” a quant asks a question that data can answer: across many markets and many weeks, did the ones that rose most last month keep rising next week, and by enough to pay for trading them? The answer is a set of numbers, and the numbers decide whether the idea gets anywhere near real money.
The US Labor Department’s O*NET database describes the job in one line: “Develop quantitative techniques to inform securities investing, equities investing, pricing, or valuation of financial instruments.” Its first listed task reads: “Apply mathematical or statistical techniques to address practical issues in finance, such as derivative valuation, securities trading, risk management, or financial market regulation.”
The two words overlap, but they name different halves of the work. Algorithmic trading is about execution: a program follows exact rules and places the orders. Quantitative trading is about how the rules were found and whether they deserve to be followed: the data, the hypothesis, the statistics and the test. Most quant strategies end up run by an algorithm. Plenty of algorithms were never quant: a rule copied from a forum and coded as it is has an execution layer and no evidence behind it.
Three habits make research quantitative rather than just computerized:
The method is a chain, and each link has its own way of breaking. Here it is with the question each stage answers:
| Stage | The question it answers | The usual way it breaks |
|---|---|---|
| Hypothesis | What should happen, and why would anyone be on the other side of the trade? | An idea found by staring at the data, then explained after the fact |
| Data | What did the market look like at each moment, including what later disappeared? | Only today’s survivors, or prices nobody could actually trade |
| Signal | Which number, computed from data available at the time, ranks the markets? | A number that quietly uses information from after the trade |
| Split | Which part of the history is for fitting, and which part is kept for the verdict? | Peeking at the test period while choosing the rule |
| Statistical test | Is the effect bigger than luck would produce, given how many ideas were tried? | Counting only the winning try |
| Portfolio | How much of each market, long or short, and how often to rebalance? | Sizes the market could never absorb |
| Costs | What do fees, spreads, slippage and funding take from each rebalance? | A strategy that only works at zero cost |
| Risk | What is the strategy really exposed to, and how bad can a drawdown get? | A hidden bet on the market, or a tail that never showed up in the test |
| Out-of-sample test | Does the fitted rule still work on the data it never saw? | Running it, disliking the answer, and trying again |
The mechanics of replaying a rule on past prices are a craft of their own, covered in our guide on how to backtest a trading strategy. This guide is about the layers around it: where the hypothesis comes from, how to tell an effect from luck, and how to keep the final test honest. When an AI assistant writes and runs that replay for you, our guide to AI backtesting covers the extra checks that takes.
Prices and volumes come first, usually as candles: the open, high, low and close of each period, with the amount traded. Around them sit order books and individual trades, for short-term work; derivatives data such as funding rates and open interest; company fundamentals, for stocks; and alternative data such as text or onchain flows.
For crypto perpetual futures, much of this is free. Hyperliquid’s info endpoint serves candles at intervals from one minute to one month with no key, though, in the docs’ words, “Only the most recent 5000 candles are available.” Its perpetuals endpoints return hourly funding history, 500 rows per request, and every request counts against a rate limit of 1,200 weight a minute per IP address, so a full history takes patience.
A factor is a characteristic shared by many assets that helps explain why some earn more than others: how big a company is, how cheap it looks, how it has moved lately, how much it swings with the market. Researchers sort assets by the characteristic, buy the top group, sell the bottom group, and study the spread between them. Among the best known in stocks are market beta; size and value, from Eugene Fama and Kenneth French; and momentum, from Mark Carhart. Campbell Harvey, Yan Liu and Heqing Zhu mark all of them in their catalog of factors.
Factors do two jobs. They are trading signals, and they are risk measures: commercial risk models such as MSCI’s Barra equity factor models break a portfolio’s risk and return into “style, sector and macro factors” to show what a manager is really exposed to.
Harvey, Liu and Zhu counted at least 316 factors tested in the literature by 2012, counting working papers. Because so many were tried, they argue that a newly discovered factor “needs to clear a much higher hurdle, with a t-ratio greater than 3.0,” rather than the usual 2.0, and conclude that “most claimed research findings in financial economics are likely false” (NBER working paper, 2014; Review of Financial Studies, 2016). Kewei Hou, Chen Xue and Lu Zhang then replicated 447 anomaly variables. Once tiny stocks were kept from dominating the results, 286 of them, 64%, were insignificant at the conventional 5% level, and 380, 85%, failed a t-value cutoff of three (NBER working paper, 2017).
Some do survive. Narasimhan Jegadeesh and Sheridan Titman, who documented stock momentum in 1993, went back with newer data and found that “momentum profits have continued in the 1990’s suggesting that the original results were not a product of data snooping bias” (NBER working paper, 1999; Journal of Finance, 2001). That is the real test of a factor: new data, after publication.
Yukun Liu, Aleh Tsyvinski and Xi Wu built crypto versions of the stock-market factors on 1,707 coins from 2014 to 2018. Three of them, the crypto market, size and momentum, “capture the cross-sectional expected cryptocurrency returns,” they found (NBER working paper, 2019; Journal of Finance, 2022). In the working paper, buying the coins that had risen most over the previous one to four weeks and selling those that had risen least averaged 2.5% to 4.1% a week more than one-month US Treasury bills, rebalanced weekly, while momentum measured over 8, 16, 50 or 100 weeks was not statistically significant. A published, specific, testable claim: exactly the kind of hypothesis to take to newer data.
Every backtest has a lucky version. The statistics exist to tell you how lucky yours would have to be.
Here is the whole method on real data, start to finish. The hypothesis comes from Liu, Tsyvinski and Wu: over a few weeks, crypto winners keep outperforming crypto losers. Their data ended in 2018. Hyperliquid’s perps give a fresh test: our data starts with the exchange’s first trades, in February 2023, years after their sample ended.
| Variant | Fit: return | Fit: Sharpe | Fit: t-statistic | Test: return | Test: Sharpe |
|---|---|---|---|---|---|
| 4-week momentum (chosen) | +67.2% | 1.00 | 1.34 | −14.8% | −0.17 |
| 2-week momentum | +58.8% | 0.92 | 1.24 | +7.6% | 0.32 |
| 3-week momentum | +45.9% | 0.81 | 1.09 | −15.9% | −0.18 |
| 1-week momentum | −2.2% | 0.12 | 0.16 | −18.7% | −0.37 |
| 1-week reversal | −37.8% | −0.63 | −0.85 | −12.0% | −0.18 |
| 3-week reversal | −53.4% | −1.15 | −1.55 | −13.8% | −0.12 |
| 4-week reversal | −59.5% | −1.27 | −1.71 | −12.2% | −0.10 |
| 2-week reversal | −59.5% | −1.29 | −1.73 | −32.9% | −0.68 |
Momentum made money in the fit period at three of the four lookbacks and beat its own reversal at all four, as the paper would predict. The best was 4-week momentum: +67.2% over 94 weeks with costs, a Sharpe ratio of 1.00 and a worst drawdown of 22.2%. Its information coefficient averaged 0.040, with a t-statistic of 1.68.
Look at the t-statistics before celebrating. The winner’s was 1.34: below 2 even on the data it was chosen on, and far below the 2.5 that eight tries call for. For a single rule with no real edge, luck alone would produce a t-statistic that high about 9% of the time; for the best of eight tries, far more often. A careful researcher would already call this a weak result, not a discovery.
Run once on the 71 weeks it never saw, the chosen rule returned −14.8% with costs, a Sharpe ratio of −0.17, and a worst drawdown of 40.5%, deeper than anything in the fit. It made money in 32 of 71 weeks. The signal itself turned: the information coefficient averaged −0.040 (t-statistic −1.28), so the winners of the previous four weeks did slightly worse than the losers in the week that followed. Funding, which the fit left out, softened the loss to −8.0%, for reasons the next section explains.
Every rebalance trades, and every trade pays. The 4-week momentum rule traded 91% of its book in an average week, so at 0.1% per unit traded it paid about 4.7% a year. Before costs it showed +82.0% in the fit period and −9.1% in the test; with them, +67.2% and −14.8%. Faster signals trade more: the 1-week versions traded 158% of the book a week.
Robert Novy-Marx and Mihail Velikov measured this across stock anomalies: “Most of the anomalies that we consider with one-sided monthly turnover lower than 50% continue to generate statistically significant net spreads, at least when designed to mitigate transaction costs. Few of the strategies with higher turnover do” (NBER working paper, 2014; Review of Financial Studies, 2016). A signal that has to be traded every week needs a large edge just to break even.
On a perpetual future, longs and shorts pay each other funding every hour, and Hyperliquid’s funding docs put each payment at “position_size * oracle_price * funding_rate.” A long-short book pays it on one side and collects it on the other, so the net depends on what you hold. Over the test period, the rule’s longs received funding worth 6.75% of the account in all and its shorts collected 0.53%: a net gain of 7.28%, which took the result from −14.8% to −8.0%.
The docs explain how: when a perp’s price is lower than the spot price, “the funding rate will be negative, and the short position will pay the long position.” The coins this rule bought, the ones that had risen most, paid funding in 421 of their 568 weeks in the book, as longs do at Hyperliquid’s resting rate: 7.81% of the account in all. In the other 147, their funding turned negative and they collected 14.55%, as much as 24.8% of a position’s value in a single week. The five markets that paid the rule the most added 8.90%, more than the whole net. Income that rests on a few extreme weeks in a handful of markets can vanish at the next rebalance. Many backtests leave funding out; for any perp strategy that holds for days, put it in, whichever way it points.
A risk model asks what a strategy is really exposed to. This one was built to be market-neutral, half long and half short, and it was: its weekly returns had a correlation of −0.03 with Bitcoin’s in the fit period and −0.02 in the test. Neutral to the market is not the same as safe. It still fell 40.5% from its peak in the test, because its real exposure was to a single bet, momentum, and that bet went wrong. Size positions so a drawdown twice the worst you have seen is survivable, and decide in advance what drawdown ends the strategy.
O*NET’s sample job titles for the occupation run from Quantitative Analyst, Quantitative Research Analyst and Quantitative Strategy Analyst to Portfolio Manager and Investment Strategist. Asked how much education a new hire needs, 60% of its respondents said a master’s degree, 35% a bachelor’s and 5% a doctorate. Its software list includes the programming languages C#, Java, Perl and R, the statistics packages MATLAB, SAS, SPSS and Stata, and SQL for databases.
You don’t need the title to use the method. A laptop and free tools cover the research stages:
| Stage | Tool | What it does |
|---|---|---|
| Data | Hyperliquid’s info endpoint | Candles, funding, order books and markets for every perp, free and without a key |
| Data handling | pandas | An open-source Python tool for data analysis and manipulation, such as tables of prices |
| Statistics | statsmodels | Statistical models and hypothesis tests in Python, such as the regressions behind a beta |
| Factor analysis | alphalens-reloaded | A Python library for performance analysis of predictive (alpha) stock factors; first built by Quantopian, now maintained by Stefan Jansen, under the Apache 2.0 license |
| Backtesting and execution | Our tools table | Frameworks such as backtrader, vectorbt and QuantConnect LEAN, and no-code builders |
| Plain English | Ethy | Describe a rule; it backtests it, optimizes it on a training window and checks it on a held-out one |
Ethy is an AI trading agent built to work like an end-to-end trading quant: it helps you research, build, optimize and execute strategies across Hyperliquid’s markets. You describe a rule in plain English, it shows the rules it will test, and it covers several of the steps in this guide:
Know the limits. One automation trades up to 10 markets, so a 40-market long-short book like the example is research for code, not a single automation. Conditions can’t link markets (“buy ETH when BTC drops”). And Ethy’s backtests leave funding out, so for a rule that holds perps for weeks, add it yourself, as the example did. Ethy presents backtests as estimates on past data, not a promise. It is software for research and analysis and doesn’t give investment advice.
No, though they usually work together. Algorithmic trading means a program follows exact rules and places the orders. Quantitative trading is the research behind the rules: data, a hypothesis, statistical tests, costs and an out-of-sample check. A quant strategy is normally run by an algorithm, but an algorithm can run a rule that no one ever tested.
Not to use the method. Asked how much education a new hire in the occupation needs, 60% of O*NET’s respondents said a master’s degree, 35% a bachelor’s and 5% a doctorate. But the core habits are open to anyone: write the hypothesis first, split the data, count your tries and charge every cost. Free market data and open-source tools cover the rest.
The method can’t promise that. The research record is sobering: most published stock anomalies fail to replicate, and in our Hyperliquid test, a published crypto momentum signal that looked fine on almost two years of data returned −14.8% on the sixteen months that followed, −8.0% with funding. The method doesn’t make a strategy profitable; it stops you trusting one that isn’t.
It depends on how long and how many tries. A Sharpe ratio of 1 over two years has a t-statistic of about 1.4, which luck alone produces about 8% of the time; at that ratio it takes about four years to reach a t-statistic of 2. A high backtest Sharpe picked from many variations means less than a modest one from a single rule written down in advance.
Testing a rule on data that played no part in building or tuning it. You split the history before you start, fit the rule on the first part, and run it once on the part you held back. If you change the rule after seeing that result, the held-back data stops being out of sample.
O*NET lists C#, Java, Perl and R among the occupation’s programming languages, with MATLAB, SAS, SPSS and Stata for statistics and SQL for databases. For research on daily or weekly data like our example, Python with pandas for the data and statsmodels for the statistics covers the whole method.
Yes, on Hyperliquid perps, trade.xyz’s stocks, indices and commodities, and spot tokens. Describe the rule in plain English; Ethy backtests the exact version it would run, over up to 365 days depending on the timeframe, optimizes its settings on a training window and checks each candidate on a held-out test window. Its backtests leave funding out.
Trading involves risk, including the loss of your money. Ethy AI is a research and trading tool and does not give investment advice or recommendations. Backtests are simulations on historical data and do not predict future results.
Deploy your AI trading agent in one click.
