Page 186 -

9 Verifying and Validating Simulations 183

future states of the target system—and the predictions prove satisfactory in repeated
tested events, it may be reasonable to expect the model outcomes to stay reliable
under similar conditions (Gross and Strand 2000). The purpose of prediction is
somewhat problematic in social simulation:
– Models of social complexity usually show nonlinear effects in which the
global behaviour of the model can become path-dependent and self-reinforcing,
producing high sensitivity to initial conditions, which limits the use of predictive
approaches.
– Many social systems show high volatility with unpredictable events, such as
turning points of macroeconomic trade cycles or of financial markets that are
in practice (and possibly in principle) impossible to predict; refer to Moss and
Edmonds (2005) for a discussion on this.
– Many social systems are not amenable to direct observation, change too slowly,
and/or do not provide enough data to be able to compare model outcomes. Most
involve human beings and are too valuable to allow repeated intervention, which
hinders the acquisition of knowledge about its future behaviour. Policies based
on false predictions could have serious consequences, thus making the purpose
of prediction unusable (Gross and Strand 2000).
While quantitative prediction of the target system behaviour is rare or simply
unattainable, prediction in general is not able to validate per se the mechanisms of
the model as good representations of the target system. In the words of Troitzsch
(2004), “What simulations are useful to predict is only how a target system might
behave in the future qualitatively.” But a different model using different mechanisms
that could lead to the same qualitative prediction may always exist, thus providing a
different explanation for the same prediction. More often, the role of predicting
future states of the target system becomes the exploration of new patterns of
behaviour that were not identified before in the target system, whereby simulation
acquires a speculative character useful as a heuristic and learning tool. What we are
predicting is really new concepts that we had not realised as being relevant just by
looking into the target.

9.3.2.2 Validity Through Retrodiction

The difference from retrodiction to prediction is that in the former the intention is
to reproduce already observed aspects of the target system. Given the existence of a
historical record of facts from the target system, the rationale of retrodictive validity
for a predictive model is the following: If the model is able to reproduce a historical
record consistently and correctly, then the model may also be trusted for the future
(Gross and Strand 2000). However, as we have mentioned, predictive models of
social complexity are uncommon in simulation. Explanation rather than prediction
is the usual motive for retrodiction. The logic of retrodictive validity is the follow-
ing: If a model is able to consistently reproduce a record of past behaviours of the
target system, then the mechanisms that constitute the model are eligible candidates