Reversion to mean is the tendency of a number that changes over time to return to its long term average value after a period above or below that.
A typical example in an investment content would the returns on a market returning to the long term trend after a spike or a dip.
While reversion to mean can be a reasonable indicator or likely long term returns, it is not often useful as a predictive tool. While one may expect a period of exceptional (high or low) returns to end, this does not tell one when to expect it to end.
A more precise definition is more difficult to pin down. Several pages of this paper are spent discussing alternate definitions.
It is clear that mean reversion is not equally applicable to all asset classes. There is evidence of reversion in a wide range investment numbers, including interest rates, volatilities and returns on equities.
The evidence on mean reversion on stock market returns is mixed, with some evidence that it is a weak effect, while other studies have found it to be strong enough to be used to derive a strategy that out-performs the market.
If prices revert to mean, this means that returns do not follow a pure random walk, and therefore the market is not efficient.
Mean reversion in relationships between securities prices is exploited by pair trading and statistical arbitrage strategies. It is also often used to gauge whether the market as a whole is over-valued or under-valued, by metrics such as Tobin's Q cyclically adjusted PE.