Dmitry Vasilyev (VNIRO, Moscow) Sigurd Tjelmeland (IMR, Bergen)

For several years in Norwegian spring spawning herring stock assessment two different approaches were used.

One of them, the Norwegian SeaStar model, (Tjelmeland, 2004) is based on likelihoods what potentially could insure proper mutual weighting of signals from all available sources of information incorporated into the model. The second one, the Russian ISVPA (TISVPA) model, (Vasilyev, 2005, 2006; Kizner and Vasilyev, 1997) takes care about robustness of analysis and includes a number of features aiming at consistent assessment using real (that is usually noisy and containing outliers) data.

Experience of application of these models has often revealed strong differences in perceptions of the state of the stock.

There have been numerous opportunities to discuss these two approaches at ICES meetings and forums with participation of fishing industry representatives, such as the Coastal States and NEAFC meetings. These forums have repeatedly requested ICES to clarify which approach is better.

Unfortunately it was not easy to make a choice because both approaches are in fact statistically meaningful. Finally, forced by fishing industry, it was recommended to elaborate a new joint model which would include best properties of the two above mentioned approaches. Thus, in other words, just managers have promoted works at an actual and challenging scientific problem: to merge robustness, a non-classic branch of statistics, and classic statistics in terms of likelihoods, and to apply it to stock assessment modeling. As a result, two countries, Norway and Russia, agreed to develop a joint scientific project focused on this task. Two meetings of Norwegian and Russian research groups were held in 2005 and 2006 where possible approaches for general structure of a new joint model were discussed.

Thus, according to the above mentioned, the ISVPA (TISVPA) model is aimed at robustness, but what could be considered a difficulty is a need to combine various loss functions, traditional or special robust ones. The formulations of data source- specific components of the overall model loss function could be quite different depending on properties of the data set under consideration. This causes some ideological problems of their mutual weighting while their mixing into overall loss function: it could be necessary to mix, for example, the sum of squared residuals for one data source with median of squared residuals or median absolute deviation for another one, etc., that is to combine values of different statistic nature. To the contrary, implementation of likelihoods, as it is done in the SeaStar model, potentially allows to avoid this difficulty. Unfortunately, the likelihood functions, both classic and more sophisticated ones, are known to be extremely non-robust.

There are various known ways to improve the available data to make them suitable for likelihoods. But what is needed for them to be applied - to have an initial solution which is already good enough and which is obtained using available (initial) data.

Accordingly, what is really needed for herring (and not only herring) stock assessment is a model which is intrinsically robust enough to get reasonable estimates on the basis of data of real (often - dubious) quality as a starting point for further implementation of statistical procedures of data purification (ISVPA features), and statistically correct integration of information from all available sources of "improved" data by means of likelihoods (SeaStar features).

It was discussed at the two meetings of Norwegian and Russian research groups with no objections, while it was decided to go further: it was agreed to build an inclusive, flexible modeling tool for the assessment of Norwegian Spring Spawning Herring. The program should be made so that it is possible to apply it to other stocks in the future. The model is now at the stage of programming which requires a considerable amount of manpower and finances.

FISHERY IN NORTH ATLANTIC: REALITY AND PROSPECTS