ATI Learning / Trading concepts

Quantitative trading: Models, evidence and trading bots

Quantitative trading uses numerical data and mathematical or statistical models to guide trades. Understanding the method helps you check what a bot does and what its results actually establish.

Trading conceptsEvidence and executionSeptember 2026

See a worked example Check provider questions

The direct answer

What is quantitative trading?

Quantitative trading uses numerical data and mathematical or statistical models to guide trading decisions. For example, a system might buy when a measured price deviation crosses a defined threshold. Its rules can support automated or manual execution. A profitable historical test does not establish that the model will remain profitable.¹²³

The main limitation is the gap between a model and evidence that it works. A precise formula tells you how a decision happens. It does not tell you whether the signal survives trading costs, changing markets or different execution conditions.⁴⁷¹¹

ATI principle: Judge a trading model by its disclosed rules, test assumptions and supporting evidence.

Start with our guides to evaluating an automated trading system, backtest versus live results and maximum drawdown. Then use the ATI Directory to run those checks against individual system files.

From the concept to the evidence

Apply it to real systems

These four ATI profiles cover different technical environments. Open a profile to examine its documented system, performance records and remaining questions. A platform alone does not prove a quantitative edge.

Browse documented systems

Understand the label

Is quantitative trading the same as algorithmic trading?

The decision model and the way orders execute are separate things to investigate.

Quantitative and algorithmic trading overlap, but their definitions are not universal. Ernie Chan’s Wiley chapter treats them as closely related terms. Other explanations distinguish the statistical decision method from automated execution.¹

For this guide, quantitative describes the use of numbers and models to guide decisions. Algorithmic describes programmed rules or execution. That distinction helps you ask useful questions without claiming that every author uses the words identically.

A model can produce a signal that a person acts on. A programmed system can also place orders from a simple rule. Ask which decisions the software makes and which remain with the user.

Mathematical complexity is not a useful shortcut for judging a product. Research connects moving-average crossovers with time-series momentum measures. A familiar price indicator can therefore belong inside a quantitative approach.¹⁷

AI is another separate question. Research includes machine-learning methods, but quantitative trading is a broader field. Ask what a claimed AI component actually does before treating the label as meaningful evidence.²

For the broader execution concept, see what algorithmic trading means. For packaged software, see how trading bots and Expert Advisors work.

The research process

How does quantitative trading work?

A useful explanation follows the model from its hypothesis through testing and deployment.

A quantitative process turns a hypothesis into explicit rules, tests those rules and checks their behavior during operation. ATI’s framework below organizes that process into five questions. It is an assessment framework, not proof that any particular provider follows it.³¹⁹

Hypothesis

What behavior does the model seek to exploit?

Explain the premise

Data

What information was available when each decision occurred?

Check the inputs

Rules

What triggers entry, sizing and exit?

Describe the decisions

Testing

Which dates, costs and assumptions produced the results?

Separate the evidence

Operation

What happens during deployment, changes and failures?

Check the controls

Begin with the hypothesis. “The system uses mathematics” is not enough. A provider should explain the behavior it seeks to measure, even if it keeps proprietary code private.

Then inspect the data and rules. Historical testing needs information that was available at the time. Future information and a sample that excludes failed instruments can distort the result.³

Finally, ask how the provider validates and maintains the model. CFA Institute’s professional guidance emphasizes understanding model assumptions and testing outputs. Those standards apply to its members and candidates, but the questions are also useful for buyers.¹⁹

Our suggested approach is to record development dates, evaluation dates, software changes and operating controls together. That makes it easier to see whether the displayed record matches the product currently offered.

A signal you can inspect

What is a simple quantitative trading example?

A numerical signal measures a condition without proving what the market will do next.

A simple example compares a completed price with the mean and standard deviation of earlier prices. The z-score expresses the difference in standard-deviation units. NIST supports this statistical formula, not a trading threshold or prediction of recovery.²⁰

Assume a fictional system uses the previous 20 completed closes, excluding the current close. Their sample mean is 100 and their sample standard deviation is 2. The current completed close is 96.

Price-deviation z-score

z = (current close − previous-price mean) / previous-price standard deviation

With the invented inputs, z = (96 − 100) / 2 = −2.

Find the difference

96 − 100 = −4

Standardize the difference

−4 / 2 = −2

Read the result

The close is two sample standard deviations below the earlier mean.

An illustrative rule buys when the completed-bar z-score reaches −2 or lower. It submits the order after that close and accepts the next executable price. It exits when a later completed-bar score reaches zero or after five bars, whichever happens first.

These invented rules show how a model becomes a trading instruction. They are not recommended settings. A separate position limit and account-loss rule would also need to be defined before evaluating the complete process.

The order also may not fill at 96. TradingView and NinjaTrader document different assumptions for historical and real-time fills. Keep the signal price and the execution price separate when examining a result.⁷⁸

Related guide: mean reversion trading explained.

Follow the money

Can trading costs erase a quantitative edge?

Profit depends on the size of wins, losses and costs together.

Trading costs can turn a positive gross result into a negative net result. TradingView’s documentation explains why simulated performance needs cost inputs, while the CFTC includes fees and spreads among the factors buyers should examine.⁷¹⁴

Consider an invented 100-trade sample. It has 55 wins averaging $12 and 45 losses averaging $10, before costs. Each trade costs $2 for the complete entry-and-exit cycle.

Net expectancy per trade

p × W − (1 − p) × L − C

p is win frequency, W is the average gross win, L is the positive average gross-loss amount and C is average round-trip cost.

Average gross gains

0.55 × $12 = $6.60 per trade

Average gross losses

0.45 × $10 = $4.50 per trade

Subtract costs

$6.60 − $4.50 − $2 = $0.10 per trade

Check the whole sample

55 × $12 − 45 × $10 − 100 × $2 = $10

The arithmetic leaves just $10 across the sample. If round-trip costs rise to $2.50, the same trading outcomes produce a $40 loss: $660 − $450 − $250.

Count costs consistently. If reported wins and losses already include a charge, subtracting it again double-counts it. Actual bid-and-ask fills already capture spread. Subscription costs belong in the buyer’s overall expense calculation.

That is why a win rate alone cannot answer whether a system made money. In this example, the proportion of winning trades remains unchanged while the final result crosses from profit to loss.

Related guides: profit factor explained and why trading win rate can mislead.

Separate the test from the record

What can a quantitative backtest prove?

Historical tests, new-market simulations and real-money results answer different questions.

A backtest establishes how the modeled process performed under specified historical assumptions. It does not establish future profitability. Researchers use scenario analysis and sensitivity analysis to examine conditions beyond a single historical result.³

The selection process matters. A provider may test many configurations and display the strongest one. Bailey and colleagues explain how that search can select historical noise, leaving weaker performance outside the selection sample.⁴

Ask how many variants were tried and which dates remained separate from development. Also ask whether later results led to more changes. A historical period loses its value as an untouched check if the provider repeatedly uses it to choose settings.

Paper trading adds another type of evidence, but it still uses simulated execution. Alpaca documents exclusions such as market impact, latency-related slippage and order-queue position.¹¹

For a buyer, the practical task is to label each record correctly. Ask whether it is historical, demo or real-money evidence, then record its dates, settings and visibility. Avoid combining those results into one unexplained performance story.

Use the backtest versus live results guide for the execution gap. See backtesting, forward testing and overfitting for the research method.

Platform conventions

Does the trading platform change the evidence?

The test environment changes how historical inputs become simulated trades.

The platform changes test conventions and execution assumptions, so the same headline result may need different checks. The table summarizes documented handling. It does not rank platforms or establish the quality of a system using them.

Environment	Documented convention	What to request
MetaTrader 4	Native testing models intrabar prices. Its documented default Ask model uses the spread at the start of testing.⁶	Data source, modeling mode, spread settings and external data tools.
MetaTrader 5	Recorded-tick and generated-tick modes differ. Historical gaps can trigger generated-tick fallback.²¹	Tick coverage, selected mode, broker/server and execution-delay settings.
NinjaTrader 8	Standard historical fills use inferred intrabar paths. Strategy Analyzer also provides walk-forward tools.⁹¹⁰	Bar type, fill resolution and the testing configuration.
TradingView	Pine strategies place hypothetical orders through a broker emulator. Lower-timeframe inputs can improve historical detail.⁷	Emulator settings, costs and the route from signals to actual orders.
cTrader	cBot tests allow data, capital, commission and spread settings. Commission input uses volume units per million.²²	Input units, data source and fixed or random spread assumptions.
Broker API: Alpaca example	Paper orders do not check quantity against available quoted liquidity.¹¹	The broker’s paper/live differences and actual fill records.

More detailed inputs can improve a test without making it a live record. On MT5, even the selected profit-calculation mode matters. Its pips mode eliminates swap and commission calculations.⁵

Our recommended check is to request the report settings alongside the report. A screenshot without its configuration leaves important parts of the evidence unexplained.

Verification and limits

How do you verify quantitative trading results?

A verification badge supports a defined check rather than every claim made about a product.

Verify what each evidence source actually checks before accepting the provider’s description of it. Myfxbook and FX Blue publish their verification scopes. Neither scope establishes future returns or proves that one unchanged bot caused every trade.¹²¹³

On Myfxbook, track-record verification concerns matching account history with broker-platform history. Trading-privileges verification concerns trading access. FX Blue distinguishes broker-sourced verified records, publisher-based semi-verified records and unverified uploads.¹²¹³

Request the direct account link, account type, update date and available history. Then ask which software version and settings produced it. Treat the link and the provider’s explanation as separate parts of the check.

Registration and identity need a separate review. The CFTC and the joint SEC, NASAA and FINRA alert warn about purported AI schemes and guaranteed returns. A quantitative or AI label does not replace those checks.¹⁴¹⁵

The FCA advises checking authorization, relevant permissions and official contact details. It also says that absence from a warning list does not establish legitimacy.¹⁸²⁴

NFA’s hypothetical-performance guidance applies to specified members and associates. It requires prominent disclosures and material assumptions. ESMA separately warns about unreliable AI information, privacy risks and exaggerated return claims.²³¹⁶

Keep the service and jurisdiction in view. These resources do not establish that every software seller has the same registration obligations.

Related guides: trading bot scam red flags and how to verify trading bot performance.

Fictional worked example

A worked example: Check the costs before accepting the result

Two reports can show the same gross trades while supporting different conclusions.

System North and System South are fictional reports for the same invented 100 trades. They illustrate a calculation and an evidence gap, not the results of any listed system.

Trading outcomes

55 wins averaging $12 and 45 losses averaging $10, before costs.

System North

Reports $210 gross profit, with round-trip costs left unspecified.

System South

Uses $2.50 round-trip cost per trade and reports a $40 net loss.

Evidence basis

Invented historical sample. No demo or real-money record is supplied.

01
Reconcile the gross result
55 × $12 − 45 × $10 = $210. Both reports start with the same gross outcome.
OK
02
Include the stated costs
System South subtracts 100 × $2.50 = $250. Its net result is $210 − $250 = −$40.
Material concern
03
Identify the missing assumption
System North does not state costs, so its gross result cannot establish the net outcome.
Verification gap

The conclusion is limited but useful. System South loses money under its stated assumptions. System North needs more information before its result can answer the buyer’s question. Neither report establishes what would happen in live trading.

Inspect the public records

The same check on documented systems

Public records disclose results on different bases. Match each figure to its account, configuration and date.

The linked records below let you apply the same evidence checks to real systems. A Myfxbook account, an FX Blue page and a provider's strategy series answer different questions. Check the account or configuration behind each number before comparing results.

System	Public performance record	Disclosed drawdown	What to check next
Cypher Pros	Foundation account on Myfxbook, labeled Real, automated and MT5.²⁵	15.72% reported drawdown in ATI's September 2, 2026 snapshot.²⁶	Check the current update, verification badges, visible history and customer settings.
Vincere Trading	Provider-published portfolio series, with verification documentation offered on request.²⁷	The homepage displays both −13.35% and −10.39%.²⁸	Ask which series each figure describes, how they reconcile and what the verification report covers.
QuantVue	ATS performance page, separating backtests and provider-labeled forward results.²⁹	Omega Stack: −$15,474 forward maximum drawdown, March–August 2026; −$8,263 in its January 2024–February 2026 backtest.²⁹	These are configuration-specific dollar figures, not account-equity percentages. Check sizing, funding status and execution evidence.
GemAlgo	gemalgo_nikolai on FX Blue. The page exists; its official attribution remains unresolved in ATI's review.³⁰³¹	ATI's September 2 snapshot records approximately 7.7% balance drawdown and 28.7% equity valley; the tracker then showed a March 19, 2026 update.³¹	Confirm account ownership and fresh statistics. This separate forex record does not establish Orion's U.S.-equities performance.

Public source pages checked October 1, 2026. Historical snapshot dates are stated in the rows. GemAlgo's current FX Blue statistics could not be reconfirmed from the accessible page. The records use different periods, units and calculation methods; they are not a ranking or a forecast.

Use each linked ATI profile to examine the fuller evidence and remaining questions. A disclosed drawdown helps you investigate a record. It does not establish a maximum future loss or show that a customer's account will match it.

Questions before access

What should you ask a quantitative trading provider?

Ask for rules, assumptions and controls that let you connect the product with its evidence.

Ask questions that expose how the system makes decisions and how its results were produced. ATI suggests starting with the following ten questions.

01
What behavior does the model seek to exploit, and what could invalidate it?
02
Which instruments, data inputs and holding periods does it use?
03
Which decisions are automated, and which require a person?
04
What dates cover development, separate evaluation and live operation?
05
How many strategy or parameter variants were tried before selecting this one?
06
Which costs, delays and fill assumptions are included in each result?
07
What is the worst observed equity drawdown, and how was it measured?
08
What limits apply to position size, leverage, aggregate exposure and losses?
09
Which software version and settings produced the public account record?
10
Who can trade, change risk settings or move money, and how can access be revoked?

Run these questions against a real system file.

Open the ATI Directory

Also request the full fee schedule and the procedure for disconnections, rejected orders and software changes. Record unanswered questions alongside the provider’s replies. A missing answer is a gap to investigate, not a reason to invent a favorable assumption.

Common questions

Quantitative trading FAQs

The key distinctions concern the model, the evidence and the work required to use it.

Does quantitative trading require AI?

No. Mathematical trading rules can exist without machine learning. If a provider claims AI, ask which component learns and what evidence supports that claim.¹²

Can retail traders use quantitative trading?

Yes, retail platforms provide tools for testing and running programmed strategies. Access to those tools does not establish profitability. Assess the model and its evidence separately.⁵⁶¹⁰

Is quantitative trading profitable?

The category has no universal profitability result. A system’s outcome depends on its rules, payoffs, costs and execution. The fictional cost example shows why a positive gross result can become a net loss.

Is quantitative trading the same as technical analysis?

They can overlap. Price-based indicators can become quantitative rules, and research connects moving-average and momentum trend measures. Treat the actual decision process as the useful comparison.¹⁷

Do you need coding for quantitative trading?

Developing a strategy and using packaged software are different tasks. Buyers can use prebuilt platform software, but still need to understand its settings and limitations. A provider’s interface does not remove the need to examine its evidence.⁶¹⁹

What are the main quantitative trading risks?

Important risks include selection bias, flawed inputs and differences between simulated and live execution. Investigate the instrument, model and account arrangements rather than assuming the quantitative label explains the risk.³⁴⁸

Does a verified account prove that a bot works?

A verified account supports only the checks defined by the tracking service. It does not establish future returns or exclusive use of one bot. Match the record with the provider’s version and settings claims.¹²¹³

The ATI principle

A trading model needs evidence you can examine

Quantitative trading gives you a method to inspect: inputs, decision rules, testing assumptions and execution. Those details help you ask more precise questions than whether a product sounds sophisticated.

Use a system file to separate what is documented from what remains a provider claim or an unanswered question. Compare evidence on its stated basis and keep the possibility of loss in view.

Browse the ATI Directory Use the full evaluation framework