A note on ‘good starting values’ in numerical optimisation

Many optimisation problems in finance and economics have multiple local optima or discontinuities in their objective functions. In such cases it is stressed that ‘good starting points are important’. We look into a particular example: calibrating a yield curve model. We find that while ‘good starting values’ suggested in the literature produce parameters that are indeed ‘good’, a simple best-of-n–restarts strategy with random starting points gives results that are never worse, but better in many cases.