Historical simulation in investment applications

On this page

Learning Outcome Statement

Describe historical simulation and explain how it can be used in investment applications.

On the exam: You may receive a current portfolio value, observed periodic returns, a tail convention, and sometimes a changed input. You may need to describe the replay, compute or identify the empirical-tail loss, and explain how the result informs investment use. The format is usually an MCQ, a fill-in answer, or an ordered sequence.

Historical simulation takes actual observed returns and replays them on today’s portfolio. You keep today’s portfolio fixed, then ask what each old market move would have done to it. Each observed move becomes one scenario, with no parametric return distribution imposed. Sort the scenario outcomes and read the selected empirical loss tail.

Start with today’s portfolio value, $V_0$, measured in USD. Take one observed periodic return, $r_t$, written as a decimal, where $t$ identifies the period. The replayed scenario profit or loss is $P/L_t$, also measured in USD.

FeatureTreatment
Return inputObserved periodic returns
DistributionNo parametric distribution imposed
Scenario setEmpirical population
Portfolio valueCurrent value held constant
Investor signPositive gain, negative loss
Tail readingSelected empirical loss tail

Match the return period to the requested horizon before you replay it. Treat the observed history as the empirical population of scenarios, so the selected tail comes directly from sorted outcomes. Use the following equation to turn each observed return into its dollar scenario outcome.

$$ P/L_t = V_0 \times r_t $$

The multiplication replays the old return on today’s portfolio. A larger $V_0$ increases every scenario’s dollar gain or loss, while a negative $r_t$ gives the investor a loss and a positive return gives a gain. The result keeps the return’s period, and you sort every $P/L_t$ from worst loss to best gain before reading the selected empirical tail. Historical VaR is that tail loss reported as a positive amount, not a forecast guarantee.

The replay follows a fixed path.

Loading diagram…

You can compare that loss with the portfolio’s risk limit and judge whether current exposure fits the investment mandate. The method puts an observed market move behind your decision, but the tail remains a historical estimate. Read the figure as a chain: today's value stays in place while each observed return is replayed, producing dollar outcomes that can be ranked. The ranking, rather than the size of the portfolio alone, identifies which scenario supplies the selected loss tail.

Worked example

You are handed a current portfolio worth USD 1,000,000 and 10 observed daily returns. You need the one-day historical loss at confidence level 0.9, using the stated convention that the single worst observation supplies the selected tail. Use the investor perspective, so a negative scenario profit or loss is a loss. First pin down the inputs that govern the replay: the portfolio value is fixed, the daily frequency matches the one-day horizon, and the convention selects one observation from the empirical set. These details tell you what the return chart and dollar outcomes must show.

GivenValue
Portfolio valueUSD 1,000,000
Return inputDaily decimal
HorizonOne day
Confidence level0.9
Observations10
Portfolio during replayHeld constant
Additional cash flowsNone
Tail conventionSingle worst observation

The bar chart orders the most negative observed daily returns. The outcome table keeps the full empirical set for the tail decision. Read the bars as return inputs, not dollar losses: each value still has to be multiplied by USD 1,000,000. Keeping the full set matters because the selected tail is determined by ranking every replayed scenario, even though the stated convention ultimately selects just one observation.

Loading diagram…

The table is sorted from worst to best, and the observation number stays visible so you can trace the tail back to $r_8$. Compare the signed dollar outcomes down the column: negative values are investor losses, while positive values are gains. The observation label links each result to its original return, so you can verify both the arithmetic and the ordering. Since the rule selects the single worst observation, the first row supplies the tail input for the calculation that follows.

ObservationDaily returnScenario profit or loss (USD)
8-0.04-40,000
1-0.03-30,000
2-0.02-20,000
4-0.015-15,000
6-0.01-10,000
10-0.005-5,000
70.0055,000
30.0110,000
90.01515,000
50.02525,000
  1. Keep the units aligned. The daily return frequency matches the one-day horizon, so the replay creates 10 empirical one-day outcomes.

  2. Replay observation 8 first. It is the worst return in the sorted table, so apply $r_8=-0.04$ to the current USD 1,000,000 value. The substitution $P/L_8 = 1,000,000 \times (-0.04) = -40,000$ produces a scenario profit or loss of USD -40,000.

  3. Read the selected tail. With confidence level $0.9$ and 10 observations, the stated convention chooses the single worst outcome. Reverse its sign with $-(-40,000) = 40,000$, producing a historical loss of USD 40,000.

  4. Interpret the result for the investor. It is a one-day historical VaR of USD 40,000 under this replay, not a promise about the next one-day loss.

That measure lets the investor compare current exposure with a risk limit or investment mandate.

Changed condition

The current portfolio and replay convention stay fixed. Only the return in observation 8 changes.

ItemBeforeAfter
Observation 8 return-0.04-0.06
Worst scenario profit or loss (USD)-40,000-60,000
Historical VaR (USD)40,00060,000

Before looking at the changed arithmetic, predict a larger loss. The more negative return makes the scenario profit or loss more negative when it is multiplied by the same positive portfolio value, so the selected loss rises from 40,000 to 60,000. Before reading on, replay observation 8 and identify the selected loss yourself.

Common error

A candidate annualizes the daily historical returns before replaying them because VaR sounds like an annual risk measure. The shortcut feels right because annual risk figures are common in portfolio discussions. The one-day horizon paired with an annual return input exposes a unit mismatch, so use the stated daily return without annualization.

Calculator

Enter the current value and the worst return before computing the scenario. The last keystroke changes the negative scenario result into the positive loss amount.

KeysDisplay
1000000 ENTER1,000,000
0.04 +/- ENTER-0.04
×-40,000
+/-40,000

Check yourself

Practice

Which sequence describes the worked historical simulation?

Which investment use matches the historical loss produced here?

Put the historical simulation steps in order.

  • Collect observed periodic returns
  • Replay each return on current value
  • Calculate each scenario profit or loss
  • Sort the empirical outcomes
  • Read the selected loss tail
  • Compare the result with the investment mandate

Revision snapshot

Historical simulation replays observed periodic returns without imposing a parametric return distribution. Use $P/L_t = V_0 \times r_t$ for each observed period, where $V_0$ is the portfolio value at the starting point and $r_t$ is the observed return. Sort the empirical scenario outcomes before reading the selected loss tail. Keep the return frequency equal to the horizon. Do not annualize daily returns for a one-day question. Historical VaR is the loss at the selected empirical tail. It is not a forecast guarantee. On the exam, identify the replay order, compute or recognize the tail loss, and explain how an investor can compare it with a risk limit or mandate.

Unlock the full lesson and continue reading

Signup to continue reading this lesson and unlimited access questions, flashcards, AI notes, and more

Scan to download app

Scan to download
UNLOCK FREE ACCESS

Practice Questions: Historical simulation in investment applications

  1. 1. Priya Shah leads Altair Capital's project to predict rare monthly liquidity-stress events and simulate the future value of a traded asset. The prediction data contain market prices, trading volume, order-imbalance measures, and macroeconomic releases. Each observation is time-stamped according to when the information became available to an investment manager. Shah requires the research team to replicate a feasible historical decision process: a model evaluated at any date may use only information that would have been available by that date. The model-development procedure is summarized in Exhibit 1. The team separates the observations into consecutive training, validation, and final test periods. However, a junior analyst standardizes each feature using its mean and standard deviation calculated over the entire 2016–2023 sample before dividing the data into those periods. The analyst argues that standardization is harmless because no target labels are directly used. Shah notes that even unsupervised preprocessing can leak information when its parameters are estimated using future observations. The team also proposes randomly shuffled cross-validation. Shah is concerned that shuffled folds could train the model on observations occurring after observations in a validation fold. She considers either expanding-window or rolling-window validation, with each validation period occurring strictly after its associated training period. The final 2023 test set is to be used once, after model selection and hyperparameter tuning are complete. Liquidity-stress events are uncommon. Exhibit 1 reports the test-set event frequency and the initial classifier's predictions. The junior analyst emphasizes the model's 95% accuracy. Shah instead examines class-specific measures, including positive-event recall, because failing to identify an actual stress event is economically costly. Recall is defined as true positive predictions divided by the total number of actual positive events. Separately, the team simulates an asset price using the geometric Brownian motion inputs in Exhibit 2. The continuously compounded expected return and volatility are constant, and no cash distributions occur during the one-year horizon. Under the specified process, the expected terminal price is S0e^(μT), while the median terminal price is S0e^[(μ − 0.5σ²)T]. Random simulation error is assumed negligible for selecting the closest answer because 100,000 independent paths are used.

    • 2. Crestline Capital, a mid-sized asset manager specializing in equity valuation research, is developing forecasting and simulation tools using monthly financial data. Its quantitative research team first examines a market valuation ratio that appears highly persistent. The team estimates the first-order autoregressive model reported in Exhibit 1. Before interpreting coefficient t-statistics or producing forecasts, the team must determine whether the level series is covariance stationary and whether its relationship with the lagged ratio has remained stable over time. For an AR(1) process, stationarity requires the absolute value of the lag coefficient to be less than one, and the team additionally tests whether the estimated relationship differs across subperiods. The team also builds a return-prediction model from 120 standardized accounting, market, and macroeconomic predictors. Economic research suggests that only a small subset has nonzero predictive effects, but the relevant subset is not known in advance. The researchers want an initial model that controls overfitting and can perform embedded variable selection by reducing irrelevant coefficients exactly to zero. Before estimating the model, an analyst standardizes all 120 predictors using the mean and standard deviation computed over the full 2005–2024 sample. They will assess predictive performance using observations from this period. One researcher proposes ordinary random five-fold cross-validation, under which observations from all dates are randomly assigned to training and validation folds. Another researcher notes that financial prediction is chronological: when forecasting a particular month, no observation from a later month would have been available. She recommends a validation design that preserves the ordering of the observations and repeatedly estimates the model using only contemporaneously available history, expanding the estimation window one month at a time. The team separately simulates one-year return paths by resampling from historical monthly returns for a diversified multi-asset portfolio. Diagnostic work indicates economically meaningful short-run serial dependence. The current proposal resamples individual monthly observations independently with replacement. The team is considering whether to resample consecutive blocks of observations instead and is experimenting with how block length affects the trade-off between preserving dependence and maintaining path diversity. Its objective is to retain the empirical return distribution while also preserving relevant short-horizon dependence. A normal parametric simulation is available, but the researchers do not want to impose normality because the historical series displays asymmetry and heavy tails.

      • 3. Ardent Data Lab develops analytical tools for a commercial lender and an asset-management affiliate. The lender has asked the team to construct a classification model that identifies borrowers likely to default. The full data set contains completed loan outcomes, including both defaults and non-defaults. The team creates training, validation, and test samples and intends the test sample to represent previously unseen observations. Exhibit 1 describes the preprocessing and model-selection sequence. Ardent's governance policy requires that information from the final test sample not influence parameter estimation, feature engineering, hyperparameter selection, or preprocessing estimates. The lender is particularly concerned about failing to identify borrowers who will actually default. Management accepts that this priority may result in some creditworthy borrowers being classified as risky. Exhibit 2 provides the final confusion matrix. The team considers accuracy, precision, and recall but must select the metric most directly aligned with management's stated cost asymmetry. All observations belong to exactly one actual and one predicted class. The asset-management affiliate separately asks Ardent to simulate a security price under geometric Brownian motion. Exhibit 3 provides the initial price, continuously compounded expected return, volatility, and horizon. Under the assumed process, the terminal price is lognormally distributed. The expected terminal price differs from the median because the distribution is positively skewed. The analyst is specifically asked for the median, which is obtained by setting the standard normal innovation in the log-price equation equal to zero, and separately for the variance, which reflects the dispersion of the lognormal distribution rather than its central tendency. No dividends, taxes, or currency conversions apply. Finally, the affiliate evaluates a trading signal using the results in Exhibit 4. The null hypothesis is that mean monthly abnormal return equals zero, and the alternative is that it differs from zero. Statistical significance is assessed at the stated significance level. Economic significance is assessed after considering the transaction cost required to implement the signal each month. The transaction-cost estimate excludes management fees but includes expected bid–ask spreads, commissions, and market impact. The conclusions regarding the credit model, classification metric, simulated price, price variance, and trading signal must each be determined independently.

        • 4. Northlake Research develops models for a lender's investment-risk group. The first project forecasts the monthly change in a credit-market stress index. The analyst initially estimates an autoregressive model using the index level, but the estimated first-order coefficient is close to one. She performs augmented Dickey–Fuller tests on both the level and first-differenced series and reviews a Ljung–Box test of the residuals from the differenced model. The results are shown in Exhibit 1. All tests use a 5% significance level. The analyst then designs the model-validation process. Because observations are monthly and the model will predict future conditions using only information available at the forecast date, the validation process must preserve chronological order. The research director rejects any procedure that allows information from later months to enter the estimation sample used to forecast an earlier month. A final untouched test period will be retained after model selection. A second project classifies credit applications as positive or negative using a labelled data set. Exhibit 2 reports the confusion matrix from a genuinely out-of-sample period. A positive prediction means the model recommends approval. The research team evaluates precision because unnecessary positive predictions create costly manual investigations. It evaluates recall separately because missed positive cases have a different economic cost. The team compares two supervised classification models for production. A gradient-boosted tree achieves an out-of-sample area under the receiver operating characteristic curve of 0.82. A regularized logistic regression achieves an out-of-sample area of 0.76. Both results are stable across chronology-preserving validation windows and both exceed the lender's minimum required AUC of 0.74. The lender's model-governance policy requires the deployed model to provide a reproducible, directionally interpretable reason for each approval decision. Compliance staff must be able to trace the effect of each retained input directly from documented model parameters without relying on a separate approximation tool. The outcome variable is labelled, and management requires a direct estimate of approval class rather than a segmentation of similar applicants. The analyst must therefore consider predictive performance, supervision type, and interpretability together rather than selecting a model solely because it has the highest AUC.

          • 5. Atlas Investments is developing a system to forecast weekly flows and next-week returns for a highly liquid exchange-traded fund. Weekly fund flow is measured as a proportion of assets at the beginning of each week. Before estimating a forecasting equation in levels, the research team applies an augmented Dickey–Fuller test. The test's null hypothesis is that the series contains a unit root. The team uses a 5% significance level and reports the result in Exhibit 1. For an initial flow forecast, the team estimates an AR(1) model in which the current week's flow is a function of an intercept and the prior week's flow. The estimated intercept, autoregressive coefficient, and most recent observation are also reported in Exhibit 1. The one-step forecast is to be calculated directly from those estimates, with no additional trend or seasonal adjustment. A separate machine-learning classifier predicts whether the ETF's next-week return will be positive. A positive prediction causes the trading desk to increase its allocation; a negative prediction causes no trade. Because unnecessary trades incur market-impact costs, the desk is particularly interested in precision: the proportion of predicted positive observations that are actually positive. Exhibit 2 reports the classifier's results on a holdout sample containing 200 weekly observations. The raw project data include market prices, news sentiment, search activity, and ETF transactions. Every field is time-stamped, but some news variables were initially entered using the date on which the database was updated rather than the earlier publication date. The head of research instructs the team to align each variable with the date on which it first became publicly available. The team is comparing alternative validation procedures. One researcher proposes randomly reshuffling all weekly observations before each train-test split. Another proposes selecting the model using the full data set and then reporting fit on that same sample. A third proposes estimating models only on earlier observations and evaluating them on later, previously unseen observations. The investment committee wants a performance estimate that reflects genuine real-time forecasting and minimizes look-ahead and data-mining biases.

            • 6. Ardent Systematic Strategies is developing a model to forecast whether an equity index will rise or fall during the next month. The research team combines valuation spreads, macroeconomic observations, and issuer-level fundamental ratios. The investment process requires that every predictor used for a historical forecast must have been observable at the date on which that forecast would have been made. A researcher first models an aggregate valuation spread as the first-order autoregressive process reported in Exhibit 1. The innovation is white noise with a constant finite variance. Before using the spread in the forecasting system, the researcher must assess the long-run behavior implied by the estimated intercept and autoregressive coefficient. The team then compares two machine-learning models. Model M is a penalized linear regression estimated using the LASSO objective. Model N is an ensemble of sequentially fitted classification trees. Because the investment objective is directional, the target variable equals one when the next-month index return is positive and zero otherwise. The team will evaluate classification performance using measures appropriate for a binary outcome rather than treating the target as a continuous return. The current model-validation procedure randomly allocates observations from the complete history across training and validation samples. A senior researcher objects that this procedure can allow later market regimes to influence the selection of a model intended for an earlier forecast date. She proposes an expanding-window design in which each validation period occurs strictly after its corresponding training period. Hyperparameters would be selected using these validation windows, followed by a single evaluation using a previously untouched test sample. The team also reviews the data controls in Exhibit 2. The existing database stores point-in-time fundamental values corresponding to their original public-release dates. A vendor offers revised historical values that correct subsequent reporting errors, but those revisions were not known to market participants on the original dates. The senior researcher warns that replacing the point-in-time records with revised values could improve backtested results without improving a deployable forecasting process.

              • 7. Aster Credit Analytics is developing a model to forecast monthly changes in an investment-grade corporate credit spread. The available database covers January 2012 through December 2022 and includes spread levels, macroeconomic variables, market liquidity measures, and sentiment extracted from issuer filings and earnings-call transcripts. The sentiment observations are derived from unstructured text, whereas spreads and macroeconomic variables are structured numerical data. Before estimating a time-series model, an analyst performs augmented Dickey–Fuller tests. The null hypothesis for each test is that the tested series has a unit root. Exhibit 1 reports the results for the spread level and its first difference. The analyst then estimates a first-order autoregressive model for monthly spread changes. All values in the model are measured in percentage points, and the forecast requested is conditional on the most recently observed spread change. Aster also compares a random forest with ridge regression. The random forest is more flexible and achieves substantially higher fit in the training sample, but the investment team is concerned that this flexibility may capture noise rather than a stable predictive relationship. Exhibit 2 reports training and test R-squared values. The team's objective is to select the model expected to perform better on new observations, not the model that describes the training sample most closely. The original project design assigns observations from 2012–2020 to training, 2021 to validation, and 2022 to final testing. Hyperparameters are selected using the validation period, and the test period is intended to remain untouched until the final model evaluation. One analyst instead standardizes every numerical predictor using the mean and standard deviation calculated from the full 2012–2022 sample before creating the three chronological subsets. The analyst also proposes repeatedly changing hyperparameters after reviewing the 2022 test results. The project leader states that preprocessing parameters should be estimated using training observations only and then applied without re-estimation to validation and test observations. Because the observations form a time series, the leader also prefers rolling or expanding-window validation to randomly shuffled cross-validation. This approach preserves temporal order and prevents later observations from supplying information to models trained as if only earlier data were available.

                • 8. Northbridge Asset Management is developing a model to forecast next-month equity-index returns. The research team first examines the time-series properties of the index's log price level. It estimates an autoregressive model and conducts augmented Dickey–Fuller tests on both the level and its first difference. The ADF test's null hypothesis is that the tested series has a unit root. The team uses a 5% significance level and assumes that the test specifications include the appropriate deterministic terms and lag structure. The team then compares a linear prediction model with a gradient-boosted tree model. Prediction error is measured using root mean squared error, so lower values indicate better performance. Exhibit 1 reports training and validation results. The model-selection objective is out-of-sample forecasting rather than explanation of historical observations. The team recognizes that a flexible model can fit noise in the training sample and therefore produce substantially worse results on unseen observations. Exhibit 2 summarizes the project's data-handling and validation procedures. Because observations are monthly and forecasts must be generated using only information available at each forecast date, the time ordering of the data is economically relevant. The team considers replacing the random split with a procedure that repeatedly trains on an earlier window and validates on a later window. The macroeconomic database creates an additional concern. Government agencies revise previously published values, but an investor making a historical forecast would have observed only the data vintage released at that time. The current project instead uses the most recently revised value for every historical month. The analyst also labels one sample as the final test sample. After observing its results, however, she repeatedly changes tree depth, learning rate, and the number of boosting iterations. The research director notes that this practice makes the designated test sample part of the model-development process. A new, untouched test period would therefore be required before making a final production decision.

                  Flashcards: Historical simulation in investment applications

                  1. 1. What is a Monte Carlo simulation?

                    Show answer

                    The repeated generation of risk factors using assumed probability distributions to generate a distribution of security values.

                  2. 2. What is a major limitation of Monte Carlo simulation?

                    Show answer

                    The results are only as accurate as the assumptions about the probability distributions used.

                  3. 3. What is a major advantage of Monte Carlo simulation over historical simulation?

                    Show answer

                    It can test theoretical scenarios that fall outside the range of actual historical data.

                  4. 4. What is a key limitation of bootstrap resampling in simulation?

                    Show answer

                    Its inputs are limited entirely by the distribution of actual historical outcomes.

                  5. 5. In valuation, what technique randomly selects values for key variables from probability distributions to generate a range of outcomes?

                    Show answer

                    Simulation.

                  Everything you need for CFA prep

                  Get full Oncourse access with lessons, practice questions, flashcards and AI study tools.

                  GET STARTED FOR FREE
                  75% OFFLimited time offer
                  GET 75% OFF