Quantitative research uses data, models and explicit tests to investigate financial questions. A synthetic cost ledger shows why positive performance still needs valid information, fair controls and an honest research record.
Quantitative research
Quantitative research is the systematic investigation of financial questions using data, mathematical models and explicit tests. Its practical purpose is to decide what the evidence supports: a forecast, a valuation, a risk estimate, or a proposed investment rule. In trading research, that means asking whether a strategy has an economically credible advantage after implementation costs and risks.
The output is a defensible decision. An attractive performance chart is evidence to examine, rather than a substitute for the investigation. A well-supported rejection can be a successful research result.
Turn an idea into a question that can fail
Consider a researcher who suspects that temporary selling pressure causes liquid shares to rebound. “Prices bounce” is too loose to test. The research specification must say which shares qualify, how selling pressure is measured, when the measurement becomes available, what return interval is forecast, and what comparison would count against the idea.
The economic explanation helps choose meaningful controls. A rebound could reflect broad market recovery, compensation for overnight risk, or an execution convention that assumes an unavailable price. A simple portfolio with comparable exposures is a useful baseline. A predeclared comparison in an interval where the proposed mechanism should be absent can challenge the timing story. Neither control, by itself, establishes causality.
Write down the target and primary evaluation measure before inspecting final results. A model that forecasts returns accurately may still fail to produce a profitable portfolio after sizing, turnover and transaction costs. Research can also evaluate pricing or risk models without proposing a trade at all.
For a worked method of specifying the mechanism, comparator and rejection conditions, see Economic Mechanisms and Falsifiable Hypotheses (public preview; full lesson requires membership).
Make the historical information set believable
A financial dataset needs more than values and dates. The study must distinguish the period a value describes from the time it was published, received and usable by the proposed process. Revised earnings cannot silently replace the numbers available to an earlier trader. Fitting a transformation on the full dataset can create look-ahead bias even when the raw observations are correctly timestamped.
Historical eligibility matters too. Testing only today's surviving firms introduces survivorship bias. Missing values, corporate actions, unavailable borrow and excluded records need explicit treatment. A convenient data-cleaning decision can change the investment population and the conclusion.
In a backtest, those inputs must become decisions, orders, assumed fills, positions and cash movements under stated rules. The researcher should be able to trace a reported gain back through that chain.
What a positive evaluation actually says
Suppose a candidate and its evaluation protocol are frozen before an untouched evaluation set is opened. In a synthetic example, the resulting aggregate gross P&L is 20 monetary units. Total assumed executable costs are 15 units, leaving
net P&L=20−15=5 monetary units.
The cost amount represents all implementation charges included in this illustration; it is stipulated, not calibrated to a market. No capital base or elapsed duration is supplied, so the result is a monetary profit rather than a percentage or annualised return.
Synthetic evaluation ledger with gross P&L 20 monetary units, assumed executable costs 15, and net P&L 5. The remaining research questions concern information availability, a feasible baseline and the candidate's research history.
Original synthetic ledger. Equal bar scales show the gross result, cost magnitude and net remainder. A positive remainder does not settle the information, comparison or selection questions.
Holding this gross ledger fixed, costs of 20 would leave zero; costs of 25 would leave a loss of 5. These are sensitivity scenarios, not estimates of future fills. They show why an assumption near the economic break-even point deserves close attention.
Even the untouched positive result does not establish a durable edge. It supplies no sampling uncertainty, loss distribution or evidence about concentration in particular instruments or market conditions. It cannot remove selection bias from the process that generated the candidate. Its interpretation depends on how many ideas, features, filters and evaluation periods influenced the research.
Repeatedly revising a model after seeing its test results turns those results into development information. A label saying “test set” does not restore independence. Overfitting can occur across the researcher's choices as well as within a fitted model. Bailey and co-authors explain why investment backtests require attention to the search that selected the reported strategy, beyond a simple hold-out split. Their paper develops this selection problem formally.
Leave a record another researcher can challenge
Preserve the hypothesis, source data snapshot, availability rules, transformations, code version, settings, evaluation boundaries and material revisions. Retain failed trials and the reasons for exclusions. Record random seeds when methods use randomness. These artifacts allow another researcher to rerun the analysis and inspect whether its conclusion follows.
Reproducing the same output from the same inputs checks computational reproducibility. It does not show that the economic effect will persist in new data. Prospective forward testing adds new chronological evidence, while its assumed execution still needs scrutiny.
The final decision should state what failed, what remains uncertain, and what further evidence would change the recommendation. Quant Research Dossier develops this evidence package and its decision record (public preview; full lesson requires membership). The distinction matters when research moves into quantitative trading, where actual orders, capital and operational failures enter the picture.
Sources
Rishi K. Narang, Inside the Black Box, second edition (Wiley, 2013), Chapter 9, especially pp. 147–153: research, falsifiable ideas and training choices.
Rob Arnott, Campbell R. Harvey and Harry Markowitz, A Backtesting Protocol in the Era of Machine Learning, The Journal of Financial Data Science 1(1), Winter 2019, pp. 64–74, especially pp. 68–70: research motivation, documented transformations and evaluation discipline.
David H. Bailey, Jonathan M. Borwein, Marcos López de Prado and Qiji Jim Zhu, The Probability of Backtest Overfitting, author-hosted manuscript revised February 2015, Sections 1 and 5: selection among tested strategies and limitations of validation methods.
