How to Backtest a Stock Trading Strategy

How Do You Backtest a Stock Trading Strategy?

Backtesting a stock trading strategy means applying objective rules to historical data to estimate how those rules would have performed. A useful backtest tests five things at once: the logic behind the idea, the risk taken to produce the return, the consistency of results across periods, the execution the strategy would actually require, and the sensitivity of the outcome to every assumption used. Profit is one of the five.

A profitable historical result is not proof of future profitability. The purpose is narrower: to eliminate weak ideas cheaply, and to quantify how uncertain the survivors remain.

What Is a Stock-Trading Backtest?

A backtest is a historical simulation that reconstructs the trades a defined strategy would have generated, using only information that was available at each point in time.

That final clause carries most of the difficulty. Historical data sits on disk in finished form, and nothing stops you reaching forward into it:

• Future prices — filling at a price that did not yet exist.
• Current index constituents — a list that exists only because of what happened later.
• Revised financials — the restated figure, not the one originally reported.
• Surviving companies — a universe that excludes what went to zero.

None of these produce an error message. Each produces a better equity curve than reality allowed.

CFA Institute materials on quantitative strategy research identify survivorship bias, look-ahead bias, transaction-cost assumptions, data mining, and overfitting as major hazards in this kind of work — a list that reads best as a checklist, since every item can be inspected directly in your own code and data.

The five questions a complete backtest must answer

1. What causes a position to be entered?
2. What causes a position to be exited?
3. How much capital is allocated, per position and in total?
4. At what realistic price could each order have been filled?
5. How stable are the results when those assumptions change?

A test answering only the first two is a chart-reading exercise. One answering all five is a specification another person could reproduce.

What Can a Backtest Actually Tell You?

Within its assumptions, a backtest estimates:

• Trade frequency — how often the rules fired.
• Holding period — how long capital stayed committed.
• Capital exposure — how much was invested, and how often it sat idle.
• Return distribution — shape and tails, not just the average.
• Historical loss size — the worst trade and the worst cluster.
• Recovery time — how long equity spent below a previous high.
• Cost sensitivity — how much edge survives commissions, spreads, slippage.
• Period dependence — whole sample, or one favorable stretch.
• Profit concentration — how much traces to a few positions.
• Regime behavior — rising, falling, quiet, and volatile markets.
• Benchmark comparison — against a passive alternative with similar exposure.
• Practicality — whether it demands capital or attention you lack.

What it cannot do is show the range of futures. Historical markets provide one realized sequence, not every possible path — the 2010s happened once, and no strategy was tested against the other ways that decade could have unfolded.

• Indefensible: "This strategy works."
• Defensible: "Under the stated data, execution, portfolio, and validation assumptions, this strategy produced sufficiently consistent historical results to justify additional testing."

The second is longer because it names what could still be wrong.

What Makes a Backtest Credible?

Which leads to the judgment that matters most. A 40% annualized return produced with look-ahead bias is worthless — not merely optimistic, but describing an activity nobody could have performed. A moderate return with transparent logic and stable out-of-sample behavior may be far more valuable, even though it looks less impressive on a chart.

What Is the Complete Backtesting Process?

1. State the hypothesis — why the opportunity might exist.
2. Convert it into exact rules — entry, exit, timing, sizing, conflicts.
3. Define the universe — what was eligible on each historical date.
4. Acquire and validate the data — prices, corporate actions, fundamentals.
5. Establish signal timing — when each input was observable.
6. Model orders and execution — fills, commissions, spreads, slippage, capacity.
7. Define sizing and portfolio constraints — per position, and in aggregate.
8. Separate development from validation data — split before looking at either.
9. Run the simulation — on the development period only.
10. Analyze returns, risk, trades, and exposure — together, not separately.
11. Test sensitivity and robustness — parameters, costs, timing, universe.
12. Compare to benchmarks — matched for exposure.
13. Forward-test the frozen rules — on data arriving after development.
14. Document the specification and research history — including what failed.

Steps 1 through 8 are preparation. The most common shortcut in amateur research is to jump straight to step 9.

Step 1: What Is the Strategy's Hypothesis?

A hypothesis explains why an opportunity might exist. It names a mechanism that could plausibly leave money on the table:

• Investors underreact to earnings surprises, so prices keep drifting.
• Institutions forced to liquidate on a schedule push prices temporarily below fair value.
• Index reconstitution forces mechanical trading unrelated to company value.
• Stocks that have already risen attract buyers who select on recent performance.
• Volatility clusters, so a calm period is more likely followed by calm than chaos.

"Buy when RSI falls below 30" is a rule, not a hypothesis: it specifies an action without explaining why an edge should exist, and asked why 30 rather than 25, it can only point back at the backtest. The RSI indicator can be a reasonable way to measure a condition a hypothesis describes; it is not a substitute for having one.

A stronger version: large, liquid stocks occasionally fall sharply because a holder must sell quickly for reasons unrelated to the company — a redemption, a margin call, a mandate change — and because that selling is price-insensitive, it can push the price further than the news justifies, producing a partial recovery once the forced flow ends.

Use a falsifiable hypothesis

"Momentum stocks go up" cannot be tested. A version that can names all of it:

• Market: US-listed common stocks.
• Universe: largest 1,000 by market capitalization, above a stated average daily dollar volume at each rebalance date.
• Trend filter: only stocks above their 200-day moving average.
• Ranking variable: total return including dividends.
• Lookback: twelve months.
• Skip period: the most recent month excluded.
• Holding period: one month, then re-ranked.
• Comparison group: the bottom decile on the same ranking, plus a capitalization-weighted index of the universe.
• Cost boundary: the conclusion must hold after commissions, spread, and a stated slippage assumption.

That version can be wrong, which is what makes it worth testing. The vague one can only be reinterpreted.

Step 2: How Do You Convert the Idea Into Exact Rules?

Anything left unstated will still be decided — just not by you. The strategy rules guide works through the full specification.

Universe rules

Hypothetical example — for education only.

• US-listed common stocks on a primary exchange.
• Minimum raw closing price of $5 on the evaluation date.
• Minimum 20-day average daily dollar volume of $10 million.
• No stocks within two sessions of a scheduled earnings release.
• Exchange-traded funds, closed-end funds, and preferred shares excluded.

Entry rules

An entry rule states the condition, the instrument, the direction, and the order type.

Hypothetical example — for education only.

Enter long when the stock closes at least 8% below its 10-day average closing price, trades above its 200-day moving average, and has 20-day average dollar volume above $10 million. Submit a market order for the next session's open.

Signal timing

A daily closing price is not known until the session has ended, so a signal calculated from that close cannot normally fill at that same close. The conservative default is the next session's open. This single timing issue invalidates many amateur backtests: the signal column and the fill-price column sit on the same row of the data, and it is one of the few errors that reliably improves results.

Exit rules

• A fixed stop below entry, by price or percentage.
• A volatility-based stop scaled to recent range.
• A profit target by price, percentage, or multiple of initial risk.
• A trailing stop that ratchets up as the position gains.
• A time-based exit after a fixed number of sessions.
• A signal-based exit when the entry condition no longer holds.
• A portfolio-level exit when exposure or drawdown limits are breached.

Hypothetical example — for education only.

Exit at the first of: a close above the 10-day average closing price, a stop 1.5 times the 14-day average true range below the entry fill, or ten sessions elapsed. Stops are evaluated intraday; the others at the close, filled at the next open.

Position-sizing rules

Sizing is part of the strategy, not a reporting preference. Change the method and you change the return, the drawdown depth, the concentration, the turnover, and the risk of ruin — from an identical set of signals. The position sizing and risk-per-trade guide covers the methods in full.

Conflict rules

• Twelve stocks signal and there is room for four. Which four?
• Cash covers part of a target position. Partial, or skip?
• An entry and an exit fire for the same stock on one date.
• A stop and a target both fall inside one session's range.
• A position already exists in a stock that signals again.
• Two candidates share a sector whose cap is reached.
• A signal fires during a drawdown-triggered pause.

If the answer is not coded, the engine still makes a decision — usually through an accidental implementation detail like alphabetical ticker order or whichever branch of an if-statement comes first. That becomes part of the results without anyone choosing it.

Step 3: How Should You Define the Historical Stock Universe?

The universe is the most commonly mishandled part of a stock backtest, because the convenient version is the wrong one. The historical data guide goes deeper on reconstruction.

Avoid the current-constituent trap

Test 2005 to 2025 using today's index members, and every 2005 trade is chosen from a list assembled with twenty years of hindsight — one that excludes companies that:

• failed or went bankrupt during the period;
• declined so far they no longer qualified;
• merged into another business;
• were acquired and delisted;
• fell below the index's liquidity or size requirements;
• only entered the index recently, whose earlier history is nonetheless included.

Survivorship bias is evaluating the past using a sample shaped by survival into the present. It flatters almost every long-only equity strategy, and the flattery is invisible: the curve looks like a correct one, only better.

Use historical eligibility, not permanent eligibility

A stock may have passed a $10 million dollar-volume threshold in 2022 and failed it in 2012. Both facts are correct. Do not remove the old period because the company later grew — that imports present knowledge into a past decision. Apply each filter using the values that existed on the evaluation date.

Define the security types

A database labeled "stocks" often holds more: multiple share classes, preferred shares, depositary receipts, exchange-traded funds and notes, closed-end funds, real estate investment trusts, limited partnership units, special-purpose acquisition vehicles before and after combination, rights, warrants, units, and over-the-counter listings. These differ in liquidity, dividend and tax treatment, and borrow availability, so the specification must state which are permitted.

Step 4: What Historical Data Does a Stock Backtest Need?

Price and volume data

Daily open, high, low, close, and volume suffices when decisions are made once per session and exits do not depend on the path within the day. Intraday data becomes necessary for intraday triggers, for cases where a stop and a target could both be hit in one session, and whenever order size makes execution timing material.

Why daily OHLCV is sometimes not enough

Hypothetical example — for education only.

A stock records: open $50, high $55, low $48, close $53. The data does not reveal whether $55 came before or after $48. Suppose a long position carries both a stop at $49 and a target at $54: both levels fall inside the range, so both were touched, but the outcome depends entirely on sequence. Ran to $55 first, the target filled and the trade gained. Fell to $48 first, the stop filled and the trade lost. A daily bar cannot distinguish those histories, so the engine resolves it by rule rather than by fact — and whatever rule it uses becomes an assumption in the results.

Corporate actions

A complete backtest needs forward and reverse splits, cash and special dividends, stock dividends, spin-offs and split-offs, mergers and acquisitions with the terms of consideration, rights issues, ticker and exchange changes, and delistings and bankruptcies with any residual value. A missing 4-for-1 split appears as a 75% single-day loss that never happened.

Raw versus adjusted prices

Raw prices are the prices that actually traded. Adjusted prices are historical prices restated so returns are continuous through splits and, in a total-return series, dividends.

Both are needed. Adjusted series are correct for returns, moving averages, volatility, and rankings; raw series for anything depending on the traded price or share count. A robust engine may keep an adjusted series for research plus a raw series and discrete corporate-action events for simulated trading, so dividends arrive as cash on payment dates and splits change the share count on effective dates.

A rule requiring a minimum raw price of $5 should not be applied to a retrospectively adjusted series. On adjusted data the filter drifts: a stock that traded at $8 in 2010 and has since split three-for-one may appear under $3 on that date, excluding a stock that plainly qualified.

Fundamental data must be point-in-time

Financial statement data is revised. A point-in-time database preserves three things separately: the value known on a given date, the date it first became available, and each later revision as its own observation rather than a replacement. Without that, a strategy can rank on figures nobody had at the time.

Data-quality checks

• Missing, duplicated, or holiday sessions.
• Highs below lows, or an open or close outside the range.
• Zero or negative prices and volumes.
• Single-day moves suggesting an unadjusted split.
• Stale repeated closes while volume continues.
• Volume spikes orders of magnitude from the norm.
• Tickers reusing a symbol from another company.
• Securities appearing or vanishing with no corporate action.
• Timestamps in an unexpected time zone, or dates shifted by a day.

A clean-looking equity curve does not prove the underlying data is clean; it is just as likely a symptom.

Step 5: How Should Signals and Trades Be Timed?

Every decision sits on an information clock with five timestamps: observation (when the value first existed), calculation (when it could be processed into a signal), decision (when the rule produced an instruction), submission (when the order could be sent), and earliest possible fill. Collapsing any two produces look-ahead bias.

The day-T close appears there as an input and never as a fill price. A backtest using it as both describes a trade that could not have happened.

Lag externally published information correctly

For an earnings release: was it published before the open, during the session, or after the close? Did the vendor supply a release timestamp, or only a calendar date? Would a real system have processed it in time to trade that session? Was the value later revised? When precision is unavailable, use a conservative delay — assuming a figure was actionable a session earlier than it was manufactures performance that never existed.

Prevent hidden look-ahead bias

• Using tomorrow's return as an input rather than the thing predicted.
• Centering a moving average, so each value uses observations from both sides.
• Filling missing values by carrying a later observation backward.
• Normalizing using the whole test period's mean and standard deviation.
• Ranking on a metric published after the ranking date.
• Executing at a price established before the signal existed.
• Winsorizing outliers using thresholds from the full sample.

None of these live in the order-execution code, and several look like ordinary data hygiene. Look-ahead bias enters through data preparation at least as often as through trading logic.

Step 6: How Do You Model Realistic Trade Execution?

Execution assumptions are where many otherwise-sound strategies stop being profitable. The execution modeling guide covers the mechanics in depth.

Commissions and fees

Zero advertised commission is not zero cost. A trade can still carry regulatory transaction fees, exchange and clearing fees, routing charges, borrow costs on shorts, margin financing, and currency conversion on foreign listings.

Bid-ask spread

Every round trip crosses the spread at least once. A simple model buys at the ask and sells at the bid, with an estimate that varies by each stock's liquidity. A backtest filling everything at the bar close ignores the spread entirely, since the close is a single print somewhere between bid and ask. For a months-long hold that is a rounding error; for weekly turnover in less liquid names it can consume the whole edge.

Slippage

Slippage is the difference between the price expected when an order is submitted and the price at which it actually fills. It comes from movement between decision and execution, order size against available liquidity, spread widening, queue position, volatility, and opening gaps. A conservative test should not depend on favorable slippage. The liquidity and slippage guide covers how order size interacts with depth.

Fill assumptions

• Does a limit order fill because the bar touched the price, or must it trade through?
• Does a stop fill at the stop, or at the next price after a gap?
• Are partial fills possible, and how is the remainder handled?
• Can the strategy trade more shares than the bar's volume?
• Does an opening order join the auction, or the first regular-session print?
• How long does an unfilled order stay live?

The touch-versus-trade-through question alone can change limit-order results materially, because prices often touch a level without enough volume to fill everyone resting there.

Market impact and capacity

A backtest with no size constraint will happily buy a quarter of a small cap's daily volume at the closing price. Cap each order at a stated percentage of average daily volume, such as 1%; cap each position at a percentage of free-float market capitalization; charge an impact cost that grows with participation rate; and re-run at several account sizes. A result that holds at $50,000 and collapses at $5,000,000 is still useful — but only if the limit is known.

Short-selling constraints

• The shares must have been borrowable on the trade date.
• A borrow fee accrues daily, and can be large on hard-to-borrow names.
• Borrow availability can disappear, blocking new shorts.
• Existing borrows can be recalled, forcing an untimely close.
• Dividends must be paid to the lender.
• A gap above the stop produces a loss larger than the stop implied.

Step 7: How Should Position Sizing Work in a Backtest?

Hypothetical example — for education only.

Percentage of equity. Equity is $100,000, the target position is 10% of equity, whole shares only. Target amount: $100,000 × 10% = $10,000. At $80 per share: floor($10,000 ÷ $80) = floor(125) = 125 shares, committing exactly $10,000 before costs.

Hypothetical example — for education only.

Stop-based risk. Equity is $100,000 and maximum risk per trade is 0.5%, so the risk budget is $100,000 × 0.5% = $500. Entry $50, stop $48, so risk per share is $50 − $48 = $2. Share count: $500 ÷ $2 = 250 shares — a position worth 250 × $50 = $12,500, or 12.5% of equity.

That calculation does not guarantee the loss is limited to $500. A stop is an instruction, not a price floor. If the stock closes at $49 and opens the next session at $44 on unexpected news, the stop fills near $44, producing a loss of roughly 250 × $6 = $1,500 — three times the intended amount.

Portfolio-level constraints

• Maximum concurrent positions.
• Maximum total invested percentage of equity.
• Maximum exposure to one sector or industry.
• Maximum exposure to a correlated group, however defined.
• Maximum gross and net exposure for long-short strategies.
• Cash reserve requirements and margin limits.
• A drawdown threshold that reduces size or pauses entries.

Ten positions each risking 1% of equity look like ten independent bets, but if all ten answer to one driver they can behave like a single concentrated trade when it moves. The portfolio diversification guide covers that aggregation.

Step 8: How Should You Divide In-Sample and Out-of-Sample Data?

In-sample data is the historical period used to develop the strategy — to form the idea, write the rules, choose parameters, and iterate. Out-of-sample data is the period deliberately withheld from that process, used once to evaluate rules that are already final.

The out-of-sample period must stay untouched until the rules are frozen, the parameters selected, the execution assumptions defined, and the evaluation criteria established. Repeatedly checking it and modifying the strategy converts it into another in-sample period.

Preserve chronological order

A three-way split might give the earliest 60% of the history to development, the next 20% to validation, and the most recent 20% to a final held-back test, in calendar order. Randomly mixing daily observations leaks regime and temporal information: volatility clusters, so a held-out day is surrounded by training days from the same week, and trends persist, so its direction is partly encoded in its neighbors.

Walk-forward testing

1. Train or calibrate on a defined historical window.
2. Freeze the rules and parameters that window produced.
3. Test them on the next period, which the calibration never saw.
4. Move both windows forward by a fixed step.
5. Repeat, then evaluate the concatenated out-of-sample results.

CFA Institute materials describe rolling-window or walk-forward backtesting as calibrating signals on a rolling window and then tracking them through subsequent periods, closer to the sequence an actual investor faces than a single fixed split. The in-sample and out-of-sample guide covers split design.

Which Backtesting Metrics Matter?

No single metric decides whether a strategy is good. Return, risk, trade quality, capital usage, consistency, and implementation demands must be read together, because each can be improved at another's expense.

The detail lives in three focused guides. Trading performance metrics covers win rate, profit factor, payoff ratio, R-multiples, and process metrics. Risk-reward ratio covers expectancy and the win-rate tradeoff. Drawdown and volatility covers drawdown depth, duration, and asymmetric recovery.

A strategy returning 12% while invested 20% of the time has different capital characteristics from one returning 12% while continuously invested.

How Do You Read the Equity Curve?

Do not judge by ending value alone. Inspect the slope in different market environments, the depth and length of declines from previous highs, how long the curve made no progress, whether gains arrived steadily or in a few steps, where the largest contributions came from, and how it behaved when the benchmark fell hardest.

Warning patterns

• One-period dependence. Remove the best month and the whole result disappears.
• Stair-step gains with sudden collapses. Common where a strategy implicitly sells tail risk — collecting a small premium repeatedly and paying it all back occasionally.
• Long flat periods. Not fatal, but the edge may be conditional on an environment that was absent.
• Impossibly smooth growth. May indicate underestimated costs, smoothed pricing, look-ahead bias, stale marks, illiquid securities priced unrealistically, or hidden leverage.
• Abrupt performance change. Candidates include a market-structure change, a data-provider change, a regulatory change, strategy crowding, a universe expansion, or a code modification.

Break the result into subperiods before accepting the aggregate curve.

How Do You Detect Overfitting?

Backtest overfitting is tailoring a strategy so closely to historical noise that its apparent performance does not generalize.

The danger grows with the number of indicators, parameter combinations, universes, holding periods, thresholds, filters, and test periods tried. Research by Bailey and López de Prado and coauthors on backtest overfitting shows why testing many configurations and selecting the strongest historical result can produce impressive in-sample performance that fails out of sample: the winner reflects both any real effect and the largest favorable accumulation of noise in the sample.

Hypothetical example — for education only.

A developer tests a moving-average crossover across every fast length from 5 to 50 days and every slow length from 20 to 200 days — several hundred combinations. The best result pairs a 13-day and a 47-day average. Nothing about 13 and 47 suggests a market mechanism, and with several hundred draws from a noisy process, the winner may simply be the luckiest combination.

Prefer stable regions over isolated peaks

Plot performance across the parameter space rather than reporting one point. A broad plateau where every slow length from 40 to 60 days produces a similar positive result is more reassuring than a spike at 47 that fails at 46 and 48.

Record every meaningful trial

Keep a research log with one entry per meaningful variation: the hypothesis, the date, the data period, the universe, the parameters, the code version, the result, and the reason for acceptance or rejection. It is the only record of how large the search was, and therefore the only basis for discounting the best result. Presenting only the winning variation hides the selection process.

Lock the final test set

1. Decide the held-back period before development begins.
2. Remove it from the working dataset so it cannot be accessed accidentally.
3. Complete development and robustness work on the remaining data.
4. Write down the acceptance criteria in advance.
5. Freeze the rules and assumptions, and record a version identifier.
6. Run the frozen strategy on the held-back period exactly once.
7. Report the result whether or not it is favorable.
8. If the strategy is modified, treat that period as consumed and designate a new one.

The guide to avoiding overfitting covers the diagnostics in more depth.

How Should You Test Strategy Robustness?

Parameter sensitivity

Re-run with each parameter slightly above and below its chosen value. The goal is not a new optimum but whether the conclusion survives the neighborhood. If a 20-day lookback works and a 19-day lookback does not, the strategy describes an artifact of one window length in one sample.

Cost sensitivity

Run four scenarios and report all: optimistic (best plausible fills and minimum fees, an upper bound only), expected (what you actually believe), conservative (higher costs, unfavorable fills), and stress (a difficult liquidity environment). Where profitability ends says how much room the edge has.

Timing sensitivity

Vary the execution point — next open, next close, one full bar later. If the edge disappears after a minor delay, practical execution may be difficult.

Universe sensitivity

Test the largest 500 names instead of 1,000. Raise the liquidity floor. Exclude a sector that contributed heavily. A conclusion holding only for one precise universe definition is fragile.

Subperiod analysis

Split the history into non-overlapping subperiods and evaluate each separately. A strategy may be valid but conditional: profitable in trending markets, flat in range-bound ones. A regime-dependent strategy needs explicit rules for when it is active, based on information available at the time.

Trade-order randomization and Monte Carlo methods

Resampling the sequence of historical trades helps estimate drawdown variation, losing-streak length, recovery uncertainty, and sequence risk. It does not create new market behavior by itself, and the simulation method must match the question — resampling independent trades is inappropriate when returns are autocorrelated or positions overlap.

What Benchmark Should the Strategy Beat?

Candidates include a broad market index, an equal-weighted version of it, a sector index matching the actual holdings, a size or style index where the strategy tilts, a buy-and-hold of the strategy's own universe, a cash rate for strategies often uninvested, a blend weighted to the strategy's average exposure, and a naive version of the strategy with the clever component removed. That last is underused: if a ranking-based strategy does no better than random selection from the same filtered universe, the ranking is not the source of the returns.

Match the benchmark to exposure

A strategy invested only half the time should not be judged solely against a permanently invested index without that context. More useful:

• Return per unit of time invested.
• Return per unit of volatility.
• Return per unit of maximum drawdown.
• Performance during the benchmark's worst periods.
• The benchmark scaled to the strategy's average exposure.
• Correlation to the benchmark.

Compare more than total return

• Drawdown depth and duration.
• Volatility and the shape of the return distribution.
• Behavior in falling markets.
• Consistency across subperiods.
• Turnover and trading effort.
• Tax characteristics implied by the holding periods.
• Capacity, and whether the result scales with account size.

A strategy may earn less than the index and still be useful, if it materially reduces drawdown. Equally, it may earn more only because it uses greater leverage or accepts severe tail risk. The benchmarking guide works through exposure-matched comparison.

How Do You Move From Backtesting to Forward Testing?

Forward testing is evaluating frozen rules on data arriving after development. It can run as paper trading that records signals without submitting orders, simulated execution in a practice environment, live trading at a fraction of intended size, live trading in a ring-fenced account, or signal-only tracking where entries and exits are logged and priced but never traded.

Track expected versus actual behavior

Record, per signal: the timestamp it was generated, expected and actual entry price, expected and actual share count, expected and actual exit price and exit reason, the fees actually charged, signals not acted on and why, and any manual intervention. If real fills are consistently worse than assumed, the execution model was optimistic. Live orders also pass through transmission and brokerage processing a historical simulation may not reproduce exactly.

Do not modify the rules casually

Every change produces a new strategy version, and the accumulated evidence belongs to the old one. Record what changed, why, which result motivated it, and whether a new validation period is required. A strategy modified after three losing weeks has not been improved; it has been reset. The forward testing guide covers version control.

Backtesting Checklist

Strategy definition

• The hypothesis explains why an edge might exist.
• Entry, exit, and sizing rules allow independent reproduction.
• Signal timing states when each input was known.
• Conflict rules cover competing signals and insufficient cash.
• The universe uses historical eligibility on each date.
• Permitted security types are enforced in code.

Data integrity

• Corporate actions are complete, including delistings.
• Raw and adjusted prices are used for their correct purposes.
• Fundamental data is point-in-time, revisions stored separately.
• Checks for gaps, duplicates, impossible bars, and stale prices have run.
• The universe includes companies that failed.
• The data source, version, and retrieval date are recorded.

Execution integrity

• Commissions, regulatory fees, and per-share charges are modeled.
• The spread is applied to every round trip.
• Slippage assumptions are stated and never favorable.
• Fill rules cover limit touches, stop gaps, and partial fills.
• Order size is capped relative to traded volume.
• Borrow, fees, recall, and dividends are modeled for shorts.

Statistical integrity

• The number of variations tested is logged.
• Parameters sit in stable regions, not isolated peaks.
• The out-of-sample period was untouched until rules were frozen.
• Splits preserve chronological order.
• Results hold across subperiods, or conditionality is stated.
• Cost, timing, and universe sensitivity have been tested.
• Profit is not concentrated in one period or position.

Deployment integrity

• The specification is documented with a version identifier.
• Acceptance criteria were written before the final test.
• A forward test on new data is complete.
• Expected versus actual fills are being tracked.
• Live position sizing matches what was tested.
• A predefined condition exists for stopping the strategy.
• Capacity limits are known and the account size sits inside them.

What Backtesting Cannot Do

A backtest cannot prove future profitability. It measures one realized history under one set of assumptions, and the future is not a resample of that history.

It cannot capture every liquidity or policy environment — rates, market structure, settlement rules, tick sizes, and fees all change. It cannot account for competitor response or strategy crowding: if enough capital pursues the same signal, its returns shrink, and no dataset contains the reaction of participants who had not yet found the idea. It cannot substitute for forward testing, and it cannot make a fragile idea robust.

Results also depend on choices the developer made: the universe, the cost model, the fill rules, the sizing method, the split, the metrics reported. Two competent people can test the same idea on the same data, make different defensible choices, and disagree. That is a reason to document the assumptions, and to be more interested in which ones the conclusion depends on than in the conclusion itself.

Backtesting FAQs

Related Guides

• Writing testable strategy rules — entry, exit, timing, sizing, conflicts.
• Historical stock data — corporate actions, adjustments, point-in-time fundamentals.
• In-sample vs. out-of-sample testing — split design and walk-forward windows.
• Execution modeling — commissions, spreads, slippage, fills, capacity.
• Avoiding backtest overfitting — trial counting and parameter stability.
• Benchmarking a strategy — matching the comparison to exposure.
• Forward testing frozen rules — paper trading and expectation tracking.
• Trading performance metrics — win rate, profit factor, R-multiples.
• Position sizing and risk per trade — the sizing methods a backtest must specify.