Abstract | Volatility moves in financial markets are rare and sporadic i.e., periods of low volatility typically follow each other until the regime change due to technical and/or fundamental exogenous factors. Subsequently higher volatility tends to lead to even higher volatility and vice versa. Such dynamic systems require alternative ways of representation than a static multi-layer perceptron (MLP). The prevalent models assume L-stable distributions with independent and stable increments. This contrasts with what is observed in real world. L-stable distributions miss one of the main features of financial markets – the alternation of periods of large price changes with periods of small price changes. To correct for this deficiency another, self-affine process had been introduced – Fractional Brownian Motion (FBM). FBM does not capture fat tails or fluctuations in volatility that are unrelated to the predictability of future returns. In summary, both models have strong scale-invariance property, in which the distribution of returns over different sampling intervals are identical. This property is clearly at odds with empirical observations. The key idea in this paper is that Generative Adversarial Networks (GAN) combined with fractality of data is better suited to manage volatility time series between different shock events because it is structured to maintain a memory of older points in the time series and continuously learn from them. Episodic memory that is used in GANs maintains explicit record of past events. In order to make decision, the action is chosen that has the highest value based on the outcomes of past similar situations. There are other, older models such as time-delayed window input vector (TDNN nets) and recurrent structures such as ‘Jordan’ and ‘Elman’ but in this study we focus on a more modern way of looking at dynamic time series through semi-supervised framework. |
---|