An investor who buys a share because she expects others to want it tomorrow is not judging the company. She is judging what other investors will think, and they in turn are judging her. The guessing game reduces this problem to a single number.
The game measures how many steps of reasoning about other people a player carries out, and how that depth changes with experience. It separates two ideas that students tend to merge, namely what an equilibrium prescribes and what is a good response to the people who are in the room. A class that has studied dominance can also see iterated elimination of dominated strategies at work in its own choices.
With a multiplier of two thirds and numbers between 0 and 100, the target can never exceed 66.67, so any higher guess is weakly dominated. Once those guesses are set aside, nothing above 44.44 can be a best response, and repeating the argument without end leaves only 0, which is the unique Nash equilibrium. The image goes back to the newspaper beauty contest in Chapter 12 of Keynes (1936). Reaching 0 requires that every player is rational and also that every player is certain of the rationality of all the others, to any depth.
The level-k account, introduced by Stahl and Wilson (1995) and by Nagel (1995), replaces that requirement with a hierarchy of types. A level-0 player picks without strategic thought, with 50 as the natural average. Level 1 best responds to level 0 and chooses about 33, level 2 chooses about 22, and so on. Camerer, Ho and Chong (2004) let each type respond to a mixture of all lower types and report that an average of about one and a half steps fits a wide range of games. With a multiplier above one the same reasoning pushes choices upward, a case studied by Ho, Camerer and Weigelt (1998).
In the original experiment of Nagel (1995), the first-round mean with a multiplier of two thirds was 36.7 and the median 33. Almost half of all choices lay near 33 or 22, and no participant chose 0. Choices fell from round to round without reaching the equilibrium within four rounds. Bosch-Domènech et al. (2002) collected about 7,900 entries through three newspapers and found means between 18.9 and 25.5, with the same clusters near 33 and 22 and a third cluster at the equilibrium, a choice that was rare in laboratory and classroom samples.
Across these studies the winning number of the first round lay between 13 and 25, so that a participant who played the equilibrium lost. In later rounds the whole distribution moves down, as each player responds to the target of the round before.