Algorithmic trading is a system, not a signal
An algorithmic trading strategy uses software to make or execute trading decisions according to explicit rules. Those rules can operate in microseconds, once a day, or only a few times a year. High-frequency trading is one specialized form of algorithmic trading, not a synonym for it.
The useful mental model is a chain of separate responsibilities. Research decides what information matters and how it becomes a signal. Portfolio rules decide how much risk to take. Execution decides how to trade. Monitoring checks whether the live system still matches the intended process.
1. Data and state
Define the prices, fundamentals, economic releases, positions, cash, and other information the strategy is allowed to know at each decision time.
2. Signal generation
Convert those inputs into a forecast, ranking, condition, or target. A signal expresses an idea about exposure; it is not yet an executable order.
3. Portfolio and risk rules
Turn signals into target positions while constraining leverage, concentration, liquidity, position size, and failure states such as stale or impossible data.
4. Orders and execution
Move from current positions toward targets using executable orders while accounting for spreads, slippage, commissions, financing, borrow, impact, and missed fills.
5. Monitoring and reconciliation
Compare intended behavior with broker positions, live orders, fills, and data freshness. A production system also needs a safe way to stop when reality diverges from expectations.
Separating these layers matters because automation and investment edge are different things. Software can make a process repeatable and fast, but it cannot rescue a weak hypothesis. A promising paper signal can also disappear once realistic costs, portfolio constraints, or live execution are included.
If you only need the concise concept definition, the algorithmic trading encyclopedia entry keeps that narrower reference role.
Common systematic strategy families
Trend and momentum
Trend-following systems try to stay aligned with persistent price direction. Rules can use moving averages, breakouts, time-series momentum, relative strength, or related measures. Their main challenge is usually whipsaw when markets stop trending cleanly.
Mean reversion
Mean-reversion systems look for unusually stretched prices or relationships that may move back toward a reference level. The danger is assuming that every extreme is temporary when a genuine regime change has shifted the reference itself.
Cross-sectional ranking
A ranking strategy compares assets with one another, then allocates toward those with stronger measured characteristics. Value, quality, momentum, carry, and multi-factor portfolios can all be expressed this way when the underlying data is available point in time.
Event and fundamental rules
Some systems react to scheduled releases, filings, earnings information, valuation measures, macroeconomic conditions, or policy changes. Publication timing is critical because a backtest must not use information before investors could actually have observed it.
Statistical and machine-learning methods can sit inside any of these families. More flexible models can capture relationships that simple rules miss, but they also increase the number of ways research can fit noise. Complexity should earn its place through stronger out-of-sample evidence, not through a better-looking in-sample chart.
Backtesting without fooling yourself
A backtest applies a proposed process to historical data. Its best use is adversarial: make the strategy confront real market sequences, implementation constraints, and assumptions that can break the thesis. Historical performance is evidence about one model under one research design. It is not proof of future profitability.
Four questions deserve special attention. Could every input really have been known at the simulated decision time? Does the historical universe include securities that later disappeared? Are trading costs and fills plausible? How many hypotheses, parameters, and discarded variants were tried before the reported result was chosen?
Chronological train/test splits, walk-forward evaluation, parameter sensitivity, regime analysis, and deliberately worse execution assumptions all make it harder to mistake optimization for robustness. The separate backtesting guide goes deeper on look-ahead bias, survivorship bias, structural breaks, cost modeling, and evaluation metrics.
If you want to build the mechanics yourself, Grizzly Bulls also maintains a Node.js algorithmic trading backtester tutorial and related implementation articles.
Backtest validation is a stack, not a score
A backtest can survive one test and still fail another. Execution friction, loss sequencing, finite-sample uncertainty, research mining, serial dependence, Sharpe estimation, selection-adjusted Sharpe evidence, selection-process overfitting, validation leakage, and validation path coverage are different failure modes. Grizzly Bulls keeps them separate so one strong metric does not silently stand in for evidence the research never produced.
The ten checks below are a practical review sequence, not a universal scoring system. The evidence benchmarks that follow answer different questions: how large selection effects can become under a transparent zero-edge null, what genuinely untouched fresh data looks like after a winner has been selected, how high a selected winner must clear for a chosen family-wise null threshold, and how dependence among tested variants changes the winner distribution.
- Check 1
Execution friction
Question: Does the apparent edge survive plausible turnover and per-dollar trading costs?
Useful for catching: Strategies whose gross alpha is consumed by commissions, spreads, slippage, fees, or excessive turnover.
Does not prove: That the remaining net alpha is statistically robust or executable at larger scale.
Open the quantitative check → - Check 2
Sequence risk
Question: How likely is a specified losing streak somewhere in the trade sample, and what fixed-fraction drawdown would it create?
Useful for catching: Research that treats a high average win rate as if painful loss clusters were implausible.
Does not prove: That trades are independent or that real-world losses will follow the entered constant probability.
Open the quantitative check → - Check 3
Finite-sample win-rate uncertainty
Question: How wide is the uncertainty around the observed win rate after a finite number of trades?
Useful for catching: Headline win rates that look precise despite being estimated from limited observations.
Does not prove: That the observations are independent or that the win/loss process is stable through time.
Open the quantitative check → - Check 4
Research search multiplicity
Question: How many chances did the research process have to discover a false positive?
Useful for catching: Strategy mining where the selected winner is reported without the larger family of attempted variants.
Does not prove: That Bonferroni or Sidak corrections repair look-ahead bias, data leakage, regime instability, or hidden discarded trials.
Open the quantitative check → - Check 5
Serial dependence
Question: How much independent information remains when adjacent returns are serially correlated rather than IID?
Useful for catching: Backtests that count every return observation as independent and annualize Sharpe with square-root-of-time despite measurable serial dependence.
Does not prove: That an AR(1) model captures higher-order dependence, volatility clustering, regime changes, or the full sampling distribution of every performance statistic.
Open the quantitative check → - Check 6
Sharpe estimation uncertainty
Question: How much confidence sits behind the observed Sharpe after accounting for track-record length, skewness, and kurtosis?
Useful for catching: Sharpe ratios presented as exact scores despite finite samples and non-Normal return shapes.
Does not prove: That the strategy was not selected from many trials, that returns are serially independent, or that the return process will remain stationary.
Open the quantitative check → - Check 7
Selection-adjusted Sharpe evidence
Question: How much evidence remains behind the selected Sharpe after raising the benchmark for the broader research search and non-Normal returns?
Useful for catching: Winning Sharpe ratios judged against a zero or pre-specified benchmark even though the reported strategy was selected from many attempted alternatives.
Does not prove: That the entered trial count equals effective independent search breadth, that the search log is complete, or that data and execution assumptions are unbiased.
Open the quantitative check → - Check 8
Selection-process overfitting
Question: Across symmetric in-sample and out-of-sample splits, how often does the strategy selected as best in-sample fall at or below the median out of sample?
Useful for catching: Research processes whose in-sample winners repeatedly fail to preserve relative performance on complementary data even when a single reported winner looks impressive.
Does not prove: That the surviving strategy will profit live, that the historical matrix is unbiased, or that CSCV replaces truly new data and execution validation.
Open the quantitative check → - Check 9
Validation leakage control
Question: Which training observations must be removed when forward label windows overlap a validation fold, and how much additional embargo should separate nearby data?
Useful for catching: Time-series validation that appears chronologically separated while training labels still consume information from the validation interval or its immediate neighborhood.
Does not prove: That features are point-in-time correct, observations are independent, the embargo is universally sufficient, or genuinely unseen future data will confirm the result.
Open the quantitative check → - Check 10
Validation path coverage
Question: How many leakage-aware train/test simulations and complete out-of-sample paths does a chosen CPCV group design generate?
Useful for catching: Research that relies on one historical validation path without quantifying how many alternative structured out-of-sample paths the same grouped history can support.
Does not prove: That CPCV paths are independent observations, that more paths create more market history, or that the chosen group count and holdout design are optimal.
Open the quantitative check →
Evidence benchmarks for the research process
These are not extra pass/fail checks. They are reproducible reference cases that show why a selected in-sample winner and a pre-specified strategy should not be judged against the same baseline, why an untouched holdout matters, how search breadth changes the null Sharpe hurdle, and why correlated variants should not be counted as though every trial were fully independent.
- Benchmark 1
What can strategy search manufacture?
Question: How impressive can the best reported backtest look when every tested strategy has zero true expected return?
Reference finding: With 252 daily observations and 100 independent zero-edge trials, the median winning annualized Sharpe is about 2.48 and the chance that the winner reaches Sharpe 2 is about 90.52%.
Boundary: This is an independent Normal null benchmark. Real strategy variants are usually correlated and real returns can violate the model assumptions.
Open the benchmark study → - Benchmark 2
What happens on untouched fresh data?
Question: After selecting the luckiest zero-edge backtest, what should the chosen strategy look like on a genuinely independent holdout?
Reference finding: After selecting the best of 100 one-year zero-edge backtests, the selected strategy still has median Sharpe 0 on an independent one-year holdout; only about 0.69% of holdouts match or exceed the selected in-sample median Sharpe.
Boundary: The clean reset requires an untouched holdout. Reusing validation data for further model selection makes that data part of the research search.
Open the benchmark study → - Benchmark 3
How high should the winner clear?
Question: How high must the winning Sharpe be for an entire independent strategy search to cross a chosen family-wise false-positive threshold?
Reference finding: At 5% family-wise error, 252 daily observations and 100 independent zero-edge trials require an annualized winning Sharpe of about 3.32, versus about 1.65 for one pre-specified strategy.
Boundary: The threshold assumes independent strategy trials and a Normal zero-edge return process. It does not estimate effective trial count for correlated or adaptive research.
Open the benchmark study → - Benchmark 4
What changes when strategy variants are correlated?
Question: How does dependence among zero-edge strategy trials change the null distribution of the selected winning Sharpe?
Reference finding: With 252 observations and 100 trials, the median winning Sharpe falls from about 2.48 under independence to about 1.78 at 0.50 latent dependence and about 0.79 at 0.90 latent dependence in the one-factor copula benchmark.
Boundary: The dependence parameter is a simplified Gaussian-copula input, not a direct estimate of live return correlation or a universal effective-trial-count conversion.
Open the benchmark study →
Signal algorithms and execution algorithms solve different problems
A signal algorithm tries to decide what exposure should be held. An execution algorithm tries to complete an already-decided trade efficiently. An institution can use a VWAP-style execution algorithm without making an algorithmic forecast about whether the security itself will rise.
The evidence should match the job. A signal needs evidence that its forecast or ranking has value after costs and risk constraints. Execution needs evidence about fill quality, timing, market impact, reliability, and whether the order logic behaves safely under unusual market conditions.
What changes when a strategy goes live?
Historical simulations usually hide operational uncertainty. Live market data can arrive late. A broker can reject an order. A stock can become hard to borrow. A stop can fill far from its trigger. Positions can disagree with internal state. A model can enter a market environment that never appeared in its development sample.
Treat deployment as another research stage. Start with bounded exposure, compare expected orders with actual fills and positions, retain enough logs to reconstruct decisions, monitor data freshness, and define conditions that disable the system safely. FINRA's algorithmic-trading guidance emphasizes testing, validation, monitoring, controls, and the ability to respond when automated systems malfunction.
How Grizzly Bulls uses systematic research
Grizzly Bulls publishes educational material on how systematic trading works alongside model pages that show model-specific historical behavior and current signals. Read them differently: the educational guides explain the mechanics, while each model page documents the evidence and access available for that model.
The model lineup draws on several broad evidence families, including technical mean reversion, technical trend, volatility, macroeconomic context, and monetary-policy context. Those categories describe the kinds of inputs we discuss publicly; they do not reveal the complete proprietary mechanics of any model.