Who this is for — Anyone who wants to understand how an intuition becomes a specified, simulated and monitorable strategy; anyone who must evaluate a backtest critically; and anyone already responsible for quantitative research, execution or risk.
Systematic trading makes decisions through explicit, repeatable rules. It does not necessarily coincide with automated trading: a rule can be executed manually. Nor is it identical to quantitative trading, which emphasises data, measurements and models, or algorithmic trading, which also includes algorithms dedicated solely to execution. These areas overlap, but answer different questions.
The core of the discipline is neither a rising curve nor a bot. It is a chain of testable statements: what is hypothesised, what was knowable at each instant, which order could have been submitted, what it would have cost, how the model was selected and what evidence would change the decision. A backtest is a historical simulation conditional on these assumptions; it is not a real track record, a forecast or definitive proof of future profit.
The chapter's four phases
| Phase | Decisive question | Main pages |
|---|---|---|
| 1. Specify | was the strategy described before seeing the result? | |
| 2. Simulate | were data, clock and execution genuinely available? | |
| 3. Validate | does the result survive unused data, dependence and selection? | |
| 4. Govern | does the process work forward and know when to stop? |
The phases are not a ladder producing ever greater certainty. Each removes or makes visible a class of error. A forward test may reveal a process problem, but a favourable short result does not prove that the future distribution will match the past. Live deployment also introduces impact, operational queues, human behaviour and regime changes that a simulation can represent only in part.
First distinction: four neighbouring terms
| Term | Defining property | Example | Does not imply |
|---|---|---|---|
| Systematic | decisions determined by stated rules | monthly rebalancing under predefined criteria | code or high frequency |
| Quantitative | hypotheses expressed and evaluated using data and models | cross-sectional signal estimated across a universe | automated execution |
| Algorithmic | a software procedure decides or executes actions | algorithm that splits an order through time | presence of alpha |
| Automated | the system acts without confirmation for every event | order submission and management through an API | validity of the model |
The distinction protects against two misunderstandings. First, automating a strategy does not make it correct. Second, an execution algorithm can improve how an order is traded without generating the signal that selected the position. The SEC report on algorithmic trading documents precisely the variety of uses, benefits and risks of algorithms in modern markets.
From mechanism to strategy
The classic families — trend following, cross-sectional momentum, mean reversion, carry, relative value, market making and event-driven strategies — are categories of mechanisms, not return recipes. A concrete strategy must specify universe, observables, timestamps, transformations, signals, position construction, constraints, risk, orders, costs, benchmark and monitoring rules.
Two systems both called “momentum” may have opposite exposures and risks. One may compare instruments with one another, while another observes each instrument's own time series; one may be long-only and another long-short; one may trade daily and another monthly. The family name neither replaces the specification nor demonstrates that a historical premium can be captured after costs.
The contract of a readable backtest
A serious result allows a competent reader to reconstruct at least these elements:
- Research protocol — hypothesis, universe, period, frequency, benchmark, metrics, costs and decision criterion defined before evaluation.
- Trial ledger — successful and failed variants, modifications, date and reason for every choice; the number of attempts is part of the evidence.
- Point-in-time data — actual availability, publication lag, revisions, historical universe membership, delistings and corporate actions.
- Causal clock — timestamp of information, signal calculation, order submission and first possible fill are not confused.
- Execution economics — commissions, spread, slippage, impact, capacity, borrow, funding, roll, margin and partial fills where relevant.
- Decision separation — development, selection and final evaluation do not covertly use the same sample.
- Dependence and uncertainty — overlapping trades, autocorrelation, exposure clusters and the effective number of observations are visible.
- Reproducibility — code, configuration, versions, seeds and data snapshots or hashes can be identified.
No universal percentage determines the correct split between in-sample and out-of-sample; no fixed number of trades makes a metric reliable; no Sharpe or profit-factor value proves an edge by itself. The choices depend on horizon, dependence structure, regime heterogeneity, turnover, number of models tried and economic materiality.
Validation: what it can say
Validation can show that an implementation is consistent with its specification, that the result does not visibly depend on a single choice, that certain performance appeared in periods not used to build the model and that stated costs or scenarios do not immediately erase the effect. It can also quantify how much uncertainty remains and distinguish insufficient evidence from evidence that is incompatible with the original hypothesis and its intended use.
It cannot turn a historical relationship into an immutable law. Data snooping favours the best among many attempts; the regime may change; the market may react to deployed capacity; sources and microstructure may evolve. Methods such as the Reality Check, Deflated Sharpe Ratio, bootstrap, walk-forward tests or probability of backtest overfitting address specific problems under specific assumptions: they are not universal seals of quality. Every conclusion therefore remains conditional on the design, inputs and decision that the selected method was built to assess.
Once an out-of-sample result influences features, parameters, universe or rules, that segment has entered the research process. Continuing to call it “unseen” changes the meaning of the evidence. It is preferable to preserve a truly untouched final test or transparently state the iterative nature of the analysis.
From research to live deployment
| Environment | What it observes well | What it does not fully reproduce |
|---|---|---|
| Backtest | logic, available history and modelled scenarios | future regimes, real impact, unmodelled errors |
| Paper trading | pipeline, timing, simulated orders and daily operations | real priority, capacity, economic pressure |
| Shadow mode | theoretical signals and orders on live data without submission | actual interaction with the market and broker |
| Micro-live | fills, costs and controls with limited capital | behaviour at target scale |
| Production | actual economic process | a certain counterfactual of what would have happened elsewhere |
The transition must not be automatic. Responsibilities, limits, escalation, suspension criteria and the rollback procedure are defined first. Live monitoring separates at least four diagnoses: statistical variation compatible with the model, signal deterioration, an execution or data problem, and a regime change. A loss sequence by itself does not identify which cause is present.
Federal Reserve guidance SR 26-2 is directed at banking organisations under its supervision. It is cited here for transferable principles — purpose and intended use, validation, monitoring, documentation and effective challenge — not as a universal regulatory obligation for every trader or fund.
Suggested path
First reading: Quantitative trading → Specification → Backtest → Out of sample → Forward test.
Professional deepening: point-in-time data → leakage and survivorship → costs and capacity → time-series validation → multiple testing → sample dependence → robustness and bootstrap → metrics → live governance.
This chapter reconstructs the public and established body of knowledge on systematic trading. The Emiciclo Method is a separate, subsequent layer: it is not projected backwards onto the definitions, studies or procedures described here.
Sources
- U.S. Securities and Exchange Commission, Staff Report on Algorithmic Trading in U.S. Capital Markets — taxonomy, uses, benefits and risks of algorithmic trading; it is a staff report, not a position adopted by the Commission.
- Joseph Simonian, CFA Institute Research Foundation, Investment Model Validation: A Guide for Practitioners — practical framework for backtesting, validation, benchmarking, simulation and documentation.
- Halbert White, A Reality Check for Data Snooping — inference when a rule is selected after comparing many candidates.
- David H. Bailey et al., The Probability of Backtest Overfitting — framework for estimating the risk of selecting a historically winning but fragile configuration.
- Federal Reserve, SR 26-2 — Supervisory Guidance on Model Risk Management — current sector-specific source for model governance, validation, monitoring and documentation.