Mean,
median, and mode
The most commonly used statistical metric is the mean,
which is also referred to as the average. The mean is computed by adding
together all data points and dividing by the number of data points. Given
the example above, the mean would be (50+55+52+45+58+60+40+56+58+49=523)
divided by 10, or 52.3 customers per day. This can be considered the
number of customers to expect on a typical day.
In this example, all of the data points are near the mean. However, what
would happen if there were a few rogue data points that were very unusual,
but not erroneous? Such data points are often referred to in statistics
as outliers, and they can cause the mean to be of questionable value.
For example, what if there were two more days during which there were
no customers at all? In this case we would have the same total number
of customers, but 12 data points, for an average of 43.6 (523 divided
by 12) customers per day. If these two "bad" business days
were just a fluke, then the 43.6 average of customers a day is probably
not an adequate reflection of how many customers to expect on a typical
day of business.
In such a case, statisticians often turn to another measure known as
the median. To find the median of a set of data points,
we simply list them in order and take the value in the middle. If there
is an even number of data elements, then there will be no single value
in the middle, so we take the average of the middle two.
In the previous example with two additional days without customers, the
values listed in order are: 0, 0, 40, 45, 49, 50, 52, 55, 56, 58, 58,
60. The middle two numbers are 50 and 52, so the median is the average
of these two numbers: 51 (50+52=102 divided by 2). Note that the median
value of 51 is much closer to the original mean (before we added the
two outlying zeroes to the data set). If we really believe that the two
days of no customers are not representative of a typical 12-day span,
then the median is a better indication of what to expect on a typical
day of business than the mean.
One final measure whose importance is less obvious is called the mode.
The mode is simply the most common value in the data set. In the original
data set above, all of the numbers occur once except for 58, which occurs
twice. The mode of the original data set is therefore 58. In the case
of a tie, there are multiple modes; so, in the expanded example where
we added two days with no customers, there would be two modes: 0 and
58.
Why would we ever care about the mode? Actually, the mode is not used very often because it tends to be very close to the mean. However, for some
probability distributions you will encounter cases where the mode can be quite different from the mean. In such cases, the mode is a much easier
parameter to visualize when trying to describe a probability distribution.
Histograms
Often it is important to understand the “spread” of your
data, i.e., how much individual values tend to differ from the mean,
median, and mode. The simplest way is to create a graphical interpretation
known as a histogram. To generate a histogram, you divide the range of
data points into several smaller ranges of equal size, which are sometimes
referred to as bins. You then count the number of data points in each
range or bin. For example, the table below indicates one possible choice
of dividing up the range for the original data set above.
Graphically, a histogram is just a bar chart:
As you can see, the histogram quickly shows how spread out
the data is from the mean (52.3), median (51), and mode (58).
Populations
vs. samples
Before going further, it's important to address the distinction between
a sample and a population. The data presented above, for example, is
a sample. It contains information regarding 10 consecutive business days;
however, data for other business days is not available. The corresponding
population would be a much larger set of data, consisting of the number
of customers arriving in the store on all business days for which it
was (and ever will be) open.
Basically, the population is the set of all possible data points for
some measure, whereas a sample is some smaller subset of data that we
have knowledge about. Often the population, including for example data
for future days of business, is not available. Customer surveys are typical
examples of samples, since information about the entire population, i.e.,
all customers, is seldom available.
In such cases, we need to restrict our attention to a sample. We should
pay attention, however, to the sample size, since the bigger the sample
size is the better it will describe the population. There is a theorem
in statistics, which says that sample sizes of less than 30 should be
treated with caution. Details of this theorem go though beyond the scope
of this statistics overview. In the following, we will assume that our
sample is a good representative of the population.
Why bother with this distinction between population and sample? Because
another way to graphically present our data is as a percent of sample
in each range as shown below.
Although this graph only shows how our sample data is distributed,
there is a much wider interpretation if we assume that this data
is representative of the population. Under this assumption, you
can interpret the bar sizes as being the probabilities that the
number of customers arriving in the store on a given day will
be within the specified ranges.
For example, the graph above indicates that on any given day there is a
40% chance that between 53 and 58 customers will visit the store. Now,
this might be an unrealistic interpretation based on only 10 days worth
of data, but what if you had data for a 10-year span? A 10-year sample
might be a very good estimate of the population, in which case, this would
be a perfectly valid way to interpret the graph.
Graphs like the one above are referred to as probability distributions.
This one is labeled an estimated probability distribution because it is
based on such a small sample size (a probability distribution actually
describes a population rather than a sample). Information about probability
distributions is a key input to all of FinanceIsland's simulation tools,
so let's examine them further.
Probability distributions
Graphically speaking, a probability distribution is just a histogram with
percentages on the y-axis, rather than absolute numbers. Mathematically
speaking, a probability distribution is a function that describes how likely
it is that a measure (in the graph above, it was the number of customers
to visit a store on any given business day) will take on a particular value.
The probability distribution below is from a dice-rolling simulation in
which 5 dice were rolled together 10,000 times.
You may recognize the shape of this distribution, as it describes a vast
number of processes in nature or in business (as well as sums of dice rolls
and exam scores). It is often called the “bell curve” because
of its shape, but its technical name is the normal (or Gaussian) distribution.
Why does the distribution look like this? It's a result of the number of
different combinations of dice rolls that add to any given total. For example,
there is only one way to get a total of 5 on five dice rolls: to roll 5
ones. This is not a very common occurrence. On the other hand, there are
a lot of different ways to roll a total of 17, hence a far larger proportion
of dice rolls lead to a total of 17 and the bar at 17 is much higher.
The mean of this data set, which is a sample, happens to be 17.46. The
theoretical population mean is exactly 17.5. In this case, the mean of
the sample and the mean of the population are very close because of the
large number of data points that were studied (10,000).
The mode in our example is 17. You can find the mode of the sample just
by looking at the probability distribution, as all possible values are
listed on the x-axis. The tallest bar indicates the most common value,
or by definition, the mode. That the mode is so close to the mean is expected,
as this distribution is symmetrical, i.e., it's neither skewed to the right,
nor to the left. In general, the population mean will equal the mode for
a symmetrical distribution.
Standard deviation
Now that you know how to graphically determine the spread of your data,
we'll show you why it's worthwhile to numerically measure the spread. The
most common metric for measuring the spread is called the standard deviation,
a central concept in statistics. Most spreadsheet software and math calculators
contain functionality to calculate standard deviation. It's important to
keep in mind that it's very difficult to interpret a standard deviation
unless you know how your data is distributed.
In the case of a normal distribution, roughly 69% of all the data lie
between one standard deviation to the left of the mean and one standard
deviation to the right of the mean. For example, the standard deviation
of the data shown above (the 5 dice rolled) was calculated to be 3.83
and the mean was found to be 17.46. Since this is a normal distribution,
69% of the time the sum of the 5 dice rolls was between 17.46-3.83 and
17.46+3.83, or between 13.63 and 21.29. This really means between 14
and 21, as sums of dice rolls are always whole numbers.
Another useful rule is that roughly 95% of all the data for a normal
distribution lies within two standard deviations of the mean. In this
example, that would correspond to values between 9.8 and 25.1, or between
10 and 25 as the sums of dice rolls must be whole numbers. Below is the
normal distribution graph from above with lines inserted at various standard
deviations (SD) from the mean.
Confidence intervals
The vertical lines in the chart above are referred to as confidence limits
because they reflect how confident you can be that a random observation
of your data (in this case, a roll of 5 dice) will lead to a value within
certain bounds. The range between confidence limits is referred to as the
confidence interval. The lines shown above define two-sided confidence
intervals because they lie on either side of the mean. For example, the
range of values between the gray lines, representing two standard deviations
around the mean, is referred to as the 95% confidence interval.
One-sided confidence intervals can also be defined. For example, the 95% one-sided confidence limit for a normal distribution lies 1.64 standard
deviations from the mean. In the example above, the mean is 17.46 and the standard deviation is 3.83. This means that 95% of the area (or 95% of
all the data) lies to the left of 17.46+1.64*3.83, or to the left of 24. Alternatively, you can say that 95% of all the data will be to the right
of 17.46-1.64*3.83, or to the right of 11. Both one-sided confidence intervals are shown in the charts below.
Summary
This tutorial was a brief introduction to key statistics concepts applied
in some of FinanceIsland's advanced tools. Although this statistics
background is not required to use our tools, we hope you found this tutorial
valuable. |