The pseudopower o four explanations can be illustrated by analogy with re-
gression analysis. Given a set o fevents and a su fficiently large or rich set o f
possible explanatory factors, one can always derive postdictions or explana-
tions to any desired degree o ftightness. In regression terms, by expanding the
setofindependentvariablesonecanalwaysfindasetofpredictorswithany
desired correlation with the independent variable. The price one pays for
overfitting is, o fcourse, shrinkage, failure o fthe derived rule to work on a new
sample o fcases. The frequency and vehemence o fmethodological warnings
against overfitting suggest that correlational overkill is a bias that is quite re-
sistant to even extended professional training (for references, see Fischhoff &
Slovic, 1980).
An overfitted theory is like a suit tailored so precisely to one individual in
oneparticularposethatitwillnotfitanyoneelseoreventhatsameindividual
in the future or even in the present if new evidence about him comes to light
(e.g., i fhe lets out his breath to reveal a potbelly). A historian who had built an
airtight case accounting for all available evidence in explaining how the Bol-
sheviks won might be in a sad position were the USSR to release suppressed
documents showing that the Mensheviks were more serious adversaries than
had previously been thought. The price that investment analysts pay for over-
Figure 27.1
Two examples of cues used to identifying precursors of past shifts in stock prices: formation of re-
sistance and formation of support. However, one might argue that prior to the dramatic shifts at
their respective ends, these two patterns were essentially identical. In this light, an undulating pat-
tern neither predicts nor explains anything in these data.
630 Baruch Fischhoff