Using the formal descriptions of environmental features and human behaviour produced in the *detection* and *contexts* parts of the research program the third part will focus on predictive modelling (Van Leusen and Kamermans 2005, Kamermans 2006, Stančič et al. 2000, Stančič and Veljanovski 2000, Kammermans 2007).

Two main approaches will be followed: first, an explicative method at local level, predicting likely settlements positions based on physical topography, distance to networks, and overall territorial strategies; and second, a geostatistical method will be used at the regional level, taking into account the spatial structures of settlement distributions and related variables.

### Prediction using explicative models

The objective of explicative methods is to draw a theoretical map of settlement distributions over poorly studied areas, using the data and knowledge acquired in better known regions Environmental preferences , including distance to networks or resources and interactions between settlements, have been highlighted in this type of modelling. It is possible to build explanatory statistical models, giving the probability of finding a settlement in each location for each period. Typically binary logistic regression with a Logit model is used, combined with algorithms for the evaluation of model performance (Roubens 1982, Stančič and Kvamme 1999, Tomlin 1990). One major advantage of this kind of predictive modelling is the possibility to give each location a value between zero and one (where zero means no settlement and one means settlement), which allows the model to maintain local scale, whereas other methods require continuous variables. This in turn implies, for archaeological applications, limiting modelling to small areas where continuous data is available.

Binary logistic regression has been employed before in archaeological predictive modelling (Verhagen et al. 2005). Here, the innovation is in the focus on long term modelling and incorporating social factors. The use of high quality archaeological and environmental data, transformed into variables describing occupation patterns and including concepts of territorial strategies is also key. The deep time depth, covering a series of different land use practices in each place, and the evaluation of model performance through statistical indicators and field comparisons are also central to the modelling process. ISuch a model is simultaneously inductive and deductive.

### Prediction with geostatistical methods

Rather than estimating local probabilities, prediction with geostatistical methods produces predictive maps of settlement densities at regional scales. Taking in account, for a given period, the locations of known settlements and the areas of archaeological survey, a cellular model can be derived representing the distribution of settlement on a regular grid. In this type of model information is transformed from Boolean (presence/absence) to likelihood (expressed as a percentage probability). The probability distributions can be interpolated from well studied areas out to less well known regions. Geostatistical methods such as co-kriging base their predictions both on the spatial structure of the observed phenomenon (settlement distribution), and on more abstract models of explanatory phenomena (environmental and anthropogenic context, heritage). The result is a temporal series of predictive spatial models at regional scale, which can become the basis for land use and occupation dynamics, providing an overview of stable periods, crises and reorganizations.