Novel nonparametric approaches to stock assessment and regime shift prediction
Perretti, Charles Thomas
MetadataShow full item record
Ecosystem dynamics are often complex, nonlinear, and characterized by criticalthresholds or regime changes. Despite these difficulties, resource managers mustaccurately forecast species abundance and anticipate impending regime shifts in order toimplement sustainable management plans. In the first part of this thesis I explicitly describe a nonparametric method formultivariate forecasting which I call the MS-Map and evaluate its performance relative toa suite of parametric models. I found that, in the presence of noise, it is often possible toobtain more accurate forecasts from the MS-Map than from the model that was used togenerate the data. The inclusion of additional species yielded a large improvement forthe nonparametric MS-Map, a smaller improvement for the control model, and only aslight improvement for the alternative multi-species parametric model. When applied torockfish larval abundance data from the CalCOFI survey, the performance of the MS-Map improved when additional species were included. These results suggest that flexiblenonparametric modeling approaches should be considered for ecosystem management. In the second part of this thesis, using the three-group fishery model previouslystudied by Biggs et al. (2009), I tested a suite of statistical regime shift indicators underthe ecologically realistic conditions of high, correlated noise with short time series and arapidly changing driving variable. I found that all indicators perform poorly underrealistic conditions with the exception of the variance indicator. In contrast toexpectations from previous work, the noise spectrum did not have a strong effect onindicator performance. The amount of data used to calculate the indicator had a largeimpact on performance. Also contrary to prior work, I found that the value of the spectralratio was not a reliable indicator of an impending shift. Future research should focus ontechniques that incorporate multiple data sources simultaneously, thus reducing the timeneeded to detect an impending shift.