The efficiency of financial markets has been questioned, supposedly answered, challenged and re-questioned by academics since its inception by Fama (1965b). A new paradigm of thought emerged in Lo (2004, 2005) who introduced an alternative to the Efficient Market Hypothesis (EMH). This Adaptive Market Hypothesis (AMH) would describe efficiency as the interaction of market participants. Hence, efficiency would be cyclical, limited by the nature of said participants and the environment this interaction occurs within. This thesis aims to revisit the AMH by providing a practical means of testing its implications via neural networks – with the intent of forming definitions of the degrees of adaptivity (similar to the three forms of efficiency under the EMH). When measuring the performance of the neural network, the number of hidden nodes will be plotted against the error term. Literature shows that the optimum network lies at the point of intersection between the training and testing phases. Using a base sample period as the initial training phase, it is hypothesised that the optimal neural network will perform well up to some point in time. Afterwards, due to say an exogenous influence, the network’s performance will decrease unless retrained with data that incorporates the exogenous influence. As this process is repeated, a series of “optimum network points” will emerge. It is hypothesised that this series fluctuates over time, in line with major exogenous events (such as the internet bubble of 2000 or the global recession of 2007). A plot of these points over time should depict that efficiency (or rather performance of the neural network) is indeed cyclical.