Normal Distribution in Finance

By Ryan O'Connell, CFA, FRM Published October 26, 2026
Table of Contents

    The normal distribution in finance turns an assumed average return and volatility into estimates of how likely different outcomes are. It helps answer questions such as, “How unusual is a 2% daily loss?” But the answer depends on whether the bell curve describes the investment well enough. Understanding both the calculation and its assumptions helps you read risk estimates without mistaking a model probability for a market guarantee.

    What Is the Normal Distribution in Finance?

    The normal distribution, also called the Gaussian distribution, is a symmetric, bell-shaped probability distribution. Its mean, written as μ, locates the center. Its standard deviation, σ, controls the spread. Together, those two parameters specify the entire normal curve.[1]

    A Model of Possible Returns

    A normal return model puts most probability near the mean and progressively less probability farther away. A larger standard deviation spreads the same total probability over a wider range of gains and losses.

    Imagine sorting daily returns into narrow intervals and counting how many fall in each interval. That produces a histogram. A fitted normal curve offers a smooth model to compare with those bars. A similar-looking center is a starting point for analysis; it does not establish that the curve gives reliable probabilities in the tails.

    Probability is area under the curve. The total area is 1, or 100%. The area between two return values gives their interval probability; the area to the left of a loss threshold gives the probability of that return or worse. The curve’s height at a point is a density, rather than the probability of that exact return.[2] For the broader choice among normal, lognormal, binomial and other models, see probability distributions in finance.

    Start With Returns, Mean and Standard Deviation

    Specify what you are modeling before choosing a curve. For a daily simple price return:

    Simple Price Return
    Rt = (Pt – Pt-1) / Pt-1
    Change in closing price divided by the preceding closing price

    For a hypothetical Microsoft share, a move from $100 to $101 gives ($101 – $100) / $100 = 0.01 = 1%. A move from $100 to $99 gives ($99 – $100) / $100 = -1%. These are invented prices for illustration. The formula measures price change; a total-return analysis must also account consistently for dividends and other distributions, with appropriate split adjustments.

    The mean is the model’s average return for the chosen period. The standard deviation measures dispersion around that average. Moving the mean shifts the bell left or right; increasing standard deviation makes it wider and lower. Historical sample means and standard deviations can estimate these inputs, but neither calculation establishes normality.[1]

    Keep the return definition, horizon and units consistent. Daily simple returns, monthly simple returns and daily log returns are different inputs. Our standard deviation in finance guide covers volatility estimation and annualization; descriptive statistics in finance covers the underlying summary measures.

    Z-Score and Normal Probability Formulas

    A z-score expresses a return’s distance from the mean in standard-deviation units:

    Z-Score
    z = (x – μ) / σ
    Subtract the mean return from the return of interest, then divide by standard deviation
    • x: the return or threshold you want to evaluate.
    • μ: the assumed mean for the same return horizon.
    • σ: the assumed standard deviation, which must be positive.

    A z-score of -2 is two standard deviations below the mean; +2 is two above it. The standard normal distribution has mean 0 and standard deviation 1. Here, the notation N(μ, σ2) uses variance as its second parameter.[1][3]

    If returns follow the assumed normal model, the standard-normal cumulative distribution function, Φ, converts z into probability:

    Left Tail, Right Tail and Interval
    P(R ≤ x) = Φ((x – μ) / σ)
    Right tail: P(R > x) = 1 – Φ(z). Between a and b: Φ((b – μ) / σ) – Φ((a – μ) / σ).

    The reverse calculation finds a return threshold from a z-score: x = μ + z × σ. These relationships connect return units to probability areas.[3] You can calculate a descriptive z-score for nonnormal data too; using a normal CDF to assign it a probability requires the additional distributional assumption.

    The 68-95-99.7 Rule

    For a normal distribution, approximately 68%, 95% and 99.7% of probability lies within one, two and three standard deviations of the mean. These are central, two-sided ranges.[2]

    Normal-model probabilities and hypothetical daily return bands
    Range around mean Probability inside Probability below lower bound Return band if μ = 0.04%, σ = 1.10%
    μ ± 1σ 68.27% 15.87% -1.06% to +1.14%
    μ ± 2σ 95.45% 2.28% -2.16% to +2.24%
    μ ± 3σ 99.73% 0.135% -3.26% to +3.34%

    The rightmost column uses hypothetical daily inputs, not an estimate of historical S&P 500 performance. For example, the two-standard-deviation bounds are 0.04% – 2 × 1.10% = -2.16% and 0.04% + 2 × 1.10% = +2.24%. The one- and three-standard-deviation rows use the same operation with multipliers 1 and 3. Tail percentages follow the standard-normal CDF and are rounded.[3]

    The ranges describe model probabilities. A particular historical sample need not contain exactly those proportions, and the percentages do not carry over automatically to an asset with a different distribution.

    Example: Probability of a Daily Portfolio Loss

    Hypothetical S&P 500 Exposure

    Suppose a $100,000 portfolio has daily returns modeled as normal with μ = 0.04% and σ = 1.10%. Assume its return follows the S&P 500 exposure being modeled, ignoring fees, tracking differences and cash flows. These inputs are illustrative.

    Question: Under this model, what is the probability of losing 2.16% or more in one day?

    1. Standardize the threshold:
    z = (-2.16% – 0.04%) / 1.10% = -2.20% / 1.10% = -2.00.

    2. Find the left-tail area:
    The NIST table gives 0.47725 between 0 and +2. By symmetry, P(Z ≤ -2) = 0.5 – 0.47725 = 0.02275, or about 2.28%.[3]

    3. Translate the threshold into dollars:
    $100,000 × 0.0216 = $2,160.

    The model therefore assigns about a 2.28% probability to a one-day loss of $2,160 or more. It does not cap the loss at $2,160 or establish that the actual market probability is 2.28%.

    This is the probability logic behind normal-based risk estimates. For loss quantiles and the differences among risk methods, see Value at Risk.

    Watch: Normal Distribution in Finance Explained

    The animation follows daily returns from a histogram to probability areas and tail risk. Its sample-based estimates are separate from the explicitly hypothetical inputs used in this article’s calculations.

    What Does a Three-Sigma Loss Mean?

    Using the same assumed daily parameters, a z-score of -3 corresponds to:

    Three Standard Deviations Below the Mean
    x = 0.04% – 3 × 1.10% = -3.26%
    The threshold changes when the assumed mean, volatility or return horizon changes

    The normal probability of this return or worse is P(Z ≤ -3) = 0.00135, or about 0.135%. The probability outside both three-sigma bounds is twice that: 2 × 0.135% = 0.27%. Confusing those two probabilities doubles the estimated downside-event frequency.[3]

    If days were independent and the probability stayed fixed, the average waiting time to a downside event would be approximately 1 / 0.00135 = 741 trading days. Assuming 252 trading days per year for this illustration gives 741 / 252 = 2.94 years. That is the origin of the loose “once every three years” description.

    An Average Wait Is Not a Schedule

    The event could occur tomorrow, twice in a short span, or after a much longer wait. The calculation assumes a fixed probability and independent days. It does not say that actual stock markets produce three-sigma losses at this frequency.

    Normal Model vs. Observed Returns

    A bell-shaped center can coexist with fat tails: more probability in extreme outcomes than a fitted normal curve assigns. Financial-return research also documents volatility clustering, in which large absolute moves tend to cluster and calm periods tend to persist.[4]

    A Fixed Normal Model

    • One mean and standard deviation specify the shape.
    • Equal distances above and below the mean have equal tail areas.
    • The CDF gives a precise probability conditional on those assumptions.

    Return Data to Evaluate

    • Check whether losses and gains are distributed symmetrically.
    • Compare observed tail frequencies with the model.
    • Examine whether volatility changes across time.

    These are separate modeling issues. A normal marginal distribution alone says nothing about independence across days. Conversely, changing volatility can produce a fat-tailed pooled distribution even when shocks are assumed normal conditional on each period’s volatility. Clustering concerns the size of moves; it does not by itself predict tomorrow’s direction.[4] See skewness and kurtosis in returns for more on distribution shape.

    Why Does the Bell Curve Appear So Often?

    The classical central limit theorem says that standardized sums or averages of many independent, identically distributed contributions with finite variance approach a normal distribution as their number grows. This explains the intuition behind combining many small influences. It does not establish that actual market influences satisfy those conditions, or that collecting more daily observations makes individual daily returns normal.[9] Our central limit theorem and sampling distributions guide separates individual outcomes from sample averages.

    Black Monday: A Historical Counterexample

    October 19, 1987

    On Black Monday, the S&P 500 fell about 20.5% in one day. Federal Reserve research documents the date, decline and severe strains on market functioning.[5][6]

    Putting that historical return into the hypothetical daily model above gives z = (-20.5% – 0.04%) / 1.10% = -18.67, rounded. This is a comparison with assumed inputs, not an estimate of the volatility prevailing in October 1987.

    A normal model with those inputs places an extremely small probability on such a decline. The useful lesson is to question the model’s tail assumptions and parameters. The event’s occurrence does not make it mathematically impossible under a normal curve, and the hypothetical model does not tell us its actual market probability.

    How to Evaluate a Normal Return Model

    1. Define the variable. Choose simple or log returns, the observation frequency and whether you measure price or total return.
    2. Estimate consistent inputs. Calculate mean and standard deviation from the same return series and period. Label the sample window.
    3. Inspect the shape and tails. Compare a histogram with the candidate curve. A normal probability or Q-Q plot compares ordered observations with normal quantiles; systematic departures from a straight line suggest a poor fit.[7]
    4. Check time variation. Compare calmer and more volatile periods rather than assuming one long-run standard deviation describes every day.
    5. State the probability question. Distinguish a loss threshold, an upper-tail gain and a range before calculating its area.
    6. Test the decision against worse scenarios. Consider losses beyond the modeled threshold and how the conclusion changes with different volatility or tail assumptions.

    Use the calculator to explore the hypothetical example or your own stated assumptions. Keep every input in the same units: for the example, either enter 0.04, 1.10 and -2.16 in percentage units, or 0.0004, 0.011 and -0.0216 in decimal units.

    Try the Normal Distribution Calculator

    For specialized modeling of extreme outcomes, see extreme value theory in finance. A different model also requires evidence and assumptions; changing its name does not validate its forecasts.

    Common Mistakes

    • Treating a z-score as a probability. A z-score describes standardized distance. A probability comes from a specified distribution, such as the normal CDF.
    • Using both tails for a downside-only question. Below -3 is about 0.135% under normality; outside -3 to +3 is about 0.27%.
    • Confusing density height with probability. Use an area over a range. In a continuous normal model, an exact single value has probability zero.
    • Mixing units or horizons. Do not combine a daily threshold with annual volatility, or a percentage-unit return with decimal-unit inputs.
    • Misreading a normal table. Some tables show cumulative area to the left; the NIST table cited here shows area between zero and z. Check the heading before using an entry.[3]
    • Assuming a familiar-looking histogram proves safe tails. Examine extreme observations and time variation before using the model for a consequential risk decision.

    Limitations of the Normal Distribution in Finance

    Estimated Inputs and Model Risk

    A precise probability calculation can still rest on uncertain inputs or an unsuitable distribution. Historical estimates do not guarantee future returns, and a good fit near the center does not establish a good fit for rare losses.

    The normal distribution also extends without limit in both directions.[1] A fully paid, unlevered long-stock position cannot lose more than 100% of its value through the stock price alone. A normal model for its simple returns is therefore an approximation; the discrepancy becomes more relevant when modeled volatility or the horizon makes that lower boundary consequential.

    Finally, a one-period probability distribution does not describe an entire investment path. It cannot by itself determine the sequence of losses, recovery time or trading liquidity during a market disruption. Evaluate those questions separately from the area under a daily return curve.

    Frequently Asked Questions


    Not necessarily. Unusualness depends on the return’s distance from the mean relative to standard deviation. When the mean is small compared with daily volatility, modest losses can remain close to the center of the assumed distribution. Calculate the z-score before drawing a conclusion.


    No. Its threshold is the chosen mean minus three times standard deviation. The hypothetical inputs here produce -3.26%; a different asset, sample window or return horizon gives a different threshold. Calling a loss “three sigma” is incomplete without specifying those inputs.


    No. Plus or minus two standard deviations contains about 95.45% of a normal distribution. The central 95% range uses approximately plus or minus 1.96 standard deviations. A one-sided 95th percentile uses about +1.645, so central ranges and one-sided cutoffs must be distinguished.[3]


    They are different variables. If a stock’s log return over a horizon is assumed normal, its positive terminal price is lognormal because price is the starting price multiplied by the exponential of that log return. Neither assumption is established just by choosing the model.[8]


    Yes. The ranges are probability statements, so outcomes outside them are possible even under a correct normal model. A historical fit can also fail to describe a later period. Distinguish ordinary sampling variation from evidence that the parameters or distribution need to change.

    Sources and Calculation Notes

    Sources were checked October 26, 2026. Worked return inputs are hypothetical; Black Monday is the dated historical example. Probability calculations use the standard normal model, with rounding stated in the text.

    1. NIST/SEMATECH: Normal Distribution, parameters, standard form, support and estimation.
    2. NIST/SEMATECH: What do we mean by “Normal” data?, probability areas and empirical-rule percentages.
    3. NIST/SEMATECH: Standard Normal Cumulative Distribution Function, table conventions, tail areas and critical values.
    4. Engle and Patton: What good is a volatility model?, Quantitative Finance 1 (2001), sections 2.1 and 2.5.
    5. Carlson: A Brief History of the 1987 Stock Market Crash, Federal Reserve Board staff paper 2007-13, introduction.
    6. Barro and Liao: Rare Disaster Probability and Options-Pricing, revised Federal Reserve Board staff paper, January 2020, section 5.2.5.
    7. NIST/SEMATECH: Normal Probability Plot, assessing departures from normality.
    8. NIST/SEMATECH: Lognormal Distribution, definition and support.
    9. Chris Piech: Central Limit Theorem, independent contributions and finite-moment conditions.
    Disclaimer

    This article provides general educational information and does not constitute personalized investment advice. Statistical models depend on their assumptions, data and estimated parameters. Hypothetical calculations do not predict future returns or limit potential losses.

    Keep reading

    More in Quantitative Finance