Two averages, two questions
The arithmetic mean answers “what was the return in a typical single year?” The geometric mean answers “what steady yearly rate would have produced the same end result?” The second is the one that describes how money actually grew, and it is the same thing as the compound annual growth rate (CAGR).
Arithmetic mean = (r1 + r2 + … + rn) ÷ n
Geometric mean = ((1 + r1) × (1 + r2) × … × (1 + rn))1 ÷ n − 1
Each r is one year’s return as a decimal (12% is 0.12, a loss of 8% is −0.08) and n is the number of years. The product inside the brackets is the growth factor: what $1 became over the whole period. When only the start and end values are known, the same rate comes straight from them:
CAGR = (end value ÷ start value)1 ÷ years − 1
Volatility is measured by the standard deviation of the yearly returns, which is how far a typical year sits from the arithmetic mean:
s = √( Σ (r − mean)2 ÷ (n − 1) )
Dividing by n − 1 gives the sample standard deviation, the usual choice when the years entered are a sample of what the investment can do. Dividing by n gives the population figure, which is a little smaller; the calculator shows both.
Worked example: returns of 12%, −8%, 15%, 6% and −3%
- Arithmetic mean. (12 − 8 + 15 + 6 − 3) ÷ 5 = 22 ÷ 5 = 4.4%.
- Growth factor. 1.12 × 0.92 × 1.15 × 1.06 × 0.97 = 1.218376. Every $1 became about $1.22, a total return of 21.84%.
- Geometric mean. 1.2183761 ÷ 5 − 1 = 0.04029, or 4.029% a year.
- Check. $10,000 growing at 4.029% for five years gives $12,183.76, the actual end value. Growing at the arithmetic mean of 4.4% would give $12,402.31, which overstates the result by $218.55.
- Volatility. The squared distances from 4.4 are 57.76, 153.76, 112.36, 2.56 and 54.76, which add to 381.2. Dividing by 4 gives 95.3, and the square root is 9.76%. Dividing by 5 instead gives a population figure of 8.73%.
Why the geometric mean is always the lower one
Gains and losses are not symmetrical. A gain of 50% followed by a loss of 50% has an arithmetic mean of 0%, but $100 becomes $150 and then $75: a geometric mean of −13.40% a year. The loss is taken on a larger balance than the gain was earned on. In the same way a 20% fall needs a 25% rise to get back to level, and a 50% fall needs 100%.
The more the returns swing, the wider the gap. A useful approximation ties the three numbers together:
Geometric mean ≈ arithmetic mean − variance ÷ 2
Variance here is the population standard deviation squared, as a decimal. For the example, 0.044 − 0.08732 ÷ 2 = 0.0402, or about 4.02%, close to the exact 4.029%. The shortfall is sometimes called volatility drag. Only when every year’s return is identical do the two averages agree.
How to read the result
Use the right average for the job
To describe what an investment did over several years, or to compare two investments held for different lengths of time, use the geometric mean. Quoting the arithmetic mean for a multi-year period flatters the result, and the more volatile the investment, the more it flatters. The arithmetic mean is the fair summary of a single typical year.
CAGR hides the journey
Growing from $10,000 to $18,000 over 6 years is a CAGR of 10.29%, whatever happened in between. A smooth climb and a crash followed by a recovery give the same figure. The total return of 80% divided by 6 gives 13.33%, which is not a growth rate at all: it ignores compounding and overstates the yearly rate.
Small samples
A standard deviation from five yearly returns is a rough estimate. It can change a great deal when one more year is added, and it treats upside and downside swings alike.
Money moving in and out
These averages assume a single amount left untouched. If money was added or withdrawn along the way, the start and end values no longer measure performance alone, and a money-weighted measure such as the internal rate of return is the appropriate tool.
A total loss
A year of −100% leaves nothing to grow, so the geometric mean is −100% no matter what the other years were. A return below −100% is not possible for an unleveraged investment and is rejected.