The Shape of Financial Risk

In the previous lesson, we explored mean and standard deviation as key summary statistics of a dataset. In this lesson, we will take it a step further: using these parameters to understand whether a dataset resembles a normal distribution. If a dataset is close to normal, it should have minimal skewness or kurtosis. We will define these concepts and learn how to check for them in financial data.

A good starting point is to graph prices over time. In a typical stock chart:

• Y-axis: Price of the security (never negative for stocks; bonds may go negative)
• X-axis: Time (e.g., last two years)

This time series plot helps us see trends, highs and lows, and periods of high volatility.

Example question: Why can’t stock prices be normally distributed?

Answer: Normal distributions allow negative values, but stock prices cannot be negative.

When analyzing a price chart, we can ask:

• How high and low did the price go?
• Does the price oscillate frequently?
• Are there noticeable trends or periods of high volatility?
• Do price changes appear symmetric, or are rises and falls uneven?

While returns are not directly visible in the price chart, volatility can be inferred from the range between highs and lows over time.

We can transform price data into daily returns, where:

• Y-axis: Return (positive or negative)
• X-axis: Time

Unlike cumulative returns, daily returns show the performance of each day independently.

Important consideration: The time scale matters. Short-term (daily) returns capture market noise, while longer-term (monthly, quarterly) returns reveal broader economic trends. Higher frequency data gives more flexibility – you can aggregate daily returns to weekly or monthly returns, but not the other way around.

By collapsing the data to focus solely on returns (ignoring the time sequence), we can examine their distribution. This is key for understanding if the data is approximately normal, which has major advantages:

• A normal distribution can be fully described using just mean and standard deviation
• This allows easy estimation of probabilities, percentiles, and financial metrics for trading strategies, risk management, and portfolio optimization

While many statistical methods assume normality, financial returns often violate this assumption:

• The Central Limit Theorem suggests sums of independent, identically distributed variables tend toward normality.
• However, in financial markets, returns are neither independent nor identically distributed – behavior varies day to day, influenced by news, events, and trader behavior.

Thus, we cannot simply assume normality; instead, we test for it using properties like skewness (asymmetry) and kurtosis (tailedness).

One key property of a normal distribution is symmetry. If a dataset is normally distributed, the distribution mirrors itself around the mean. For example, in a standard normal distribution (mean 0, standard deviation 1), if you placed a mirror along the y-axis, each side would perfectly reflect the other.

However, not all financial return distributions are symmetric. When a distribution has a longer tail on the right, it is positively skewed; when the longer tail is on the left, it is negatively skewed.

Why does this matter? Negative skewness indicates a higher probability of extreme losses. If we blindly applied a symmetric model like the normal distribution, we could significantly underestimate risk. This is common in equity markets, where sudden drops in price – driven by panic selling – occur more frequently than equivalent upward jumps. Behavioral finance explains part of this: humans tend to remember losses more vividly than gains, influencing market behavior.

Visual checks for skewness:

• Histogram: Are the bars evenly distributed on either side of the mean?
• Density plot: Imagine dropping sand over each data point; does the overall shape resemble a normal distribution?

While formulas like Pearson’s skewness coefficient exist to quantify asymmetry, the key takeaway is conceptual: symmetric distributions allow standard deviation to reliably represent risk, whereas skewed distributions complicate risk assessment.

Another assumption of normality is that volatility remains constant. In a perfect normal distribution, the standard deviation is a fixed value, representing consistent variability around the mean.

But real-world data often violates this. Consider a mixture of distributions:

• 80% of the data comes from a normal distribution with mean 0 and standard deviation 1
• 20% comes from a normal distribution with mean 0 and standard deviation 4

If you calculate the overall standard deviation of this mixture, you get a single value (around 2). At first glance, it might seem like a reliable measure of volatility – but it masks the variability within the two underlying distributions.

This is where kurtosis becomes informative. Kurtosis measures the “tailedness” of a distribution:

• For a normal distribution, the standard deviation fully captures variability, and kurtosis is zero.
• For a mixture or other non-normal distributions, excess kurtosis is positive, indicating fatter tails and more extreme events than a normal distribution would suggest.

In the mixture example, the excess kurtosis is around 7 – clearly indicating a non-Gaussian distribution and highlighting that standard deviation alone underestimates the risk.

The normal (Gaussian) distribution is extremely useful because it allows us to summarize large datasets using just two parameters: the mean and the standard deviation. For example:

• About 68% of values fall within one standard deviation of the mean
• About 95.5% fall within two standard deviations
• About 99.7% fall within three standard deviations

In essence, the normal distribution acts like a “compression tool,” condensing potentially millions of data points into a form where we can easily estimate percentiles and probabilities.

But what if our data is not normally distributed? In that case, relying on the standard deviation alone can be misleading. The expected “distance” from the mean may no longer match reality. To assess this, we can measure excess kurtosis.

• Excess kurtosis ≈ 0 → Data is close to normal
• Excess kurtosis > 0 → Distribution has heavy tails, indicating higher probability of extreme events
• Excess kurtosis < 0 → Distribution has lighter tails than normal

For example, a mixture of two normal distributions, one narrow and one wide, can produce an overall excess kurtosis of 10. This reflects what we sometimes call the “volatility of volatility” in finance: variance itself is variable. Misestimating volatility can have serious consequences: percentile estimates become unreliable, and we may develop a false sense of security about risk.

The art of financial engineering is recognizing when normality is a reasonable approximation and when alternative models are required. This skill is essential for accurately pricing risk, designing trading strategies, or managing portfolios.

When returns are not perfectly normal, alternative distributions can be used to capture their characteristics, such as:

• F-distribution
• Skew normal distribution
• Log-normal distribution

We can formally test for skewness and kurtosis, and there are robust measures designed to handle asymmetry, such as:

• Rank-based methods
• Semi-variance, which focuses on negative returns while ignoring positive ones

Kurtosis tells us whether a distribution is peaked or flat, and this directly impacts risk assessment:

• Peaked distributions → lighter tails → fewer extreme events
• Flat distributions → heavier tails → higher likelihood of extreme events

For a standard normal distribution, kurtosis is approximately 3. We often use excess kurtosis (kurtosis − 3) to simplify interpretation:

• Positive excess kurtosis → heavier tails than normal
• Negative excess kurtosis → thinner tails than normal

The key takeaway: skewness and excess kurtosis indicate that a normal distribution may not adequately represent the data. If normality holds, the mean and standard deviation are reliable for estimating probabilities. If not, we must consider alternative distributions or nonparametric approaches that rely directly on the data without assuming a specific distribution.

Finally, remember that volatility reflects investor tolerance for risk. Some investors embrace larger fluctuations; others are more risk-averse. As financial engineers, our responsibility is to choose models and metrics that accurately capture this volatility, whether it behaves predictably or exhibits complex, dynamic patterns.