Using Artificial Market Models to Forecast Financial Time-Series

TL;DR

This paper introduces a systematic methodology using artificial market models and optimization techniques to forecast financial time-series, identifying predictable market segments and updating predictions dynamically.

Contribution

It presents a formal, iterative approach to detect market predictability pockets and update strategy distributions, improving upon previous preliminary methods.

Findings

Systematic identification of market predictability pockets

Effective updating of strategy estimates during market shifts

Enhanced forecasting accuracy through iterative model resets

Abstract

We discuss the theoretical machinery involved in predicting financial market movements using an artificial market model which has been trained on real financial data. This approach to market prediction - in particular, forecasting financial time-series by training a third-party or 'black box' game on the financial data itself -- was discussed by Johnson et al. in cond-mat/0105303 and cond-mat/0105258 and was based on some encouraging preliminary investigations of the dollar-yen exchange rate, various individual stocks, and stock market indices. However, the initial attempts lacked a clear formal methodology. Here we present a detailed methodology, using optimization techniques to build an estimate of the strategy distribution across the multi-trader population. In contrast to earlier attempts, we are able to present a systematic method for identifying 'pockets of predictability' in real-world markets. We find that as each pocket closes up, the black-box system needs to be 'reset' - which is equivalent to saying that the current probability estimates of the strategy allocation across the multi-trader population are no longer accurate. Instead, new probability estimates need to be obtained by iterative updating, until a new 'pocket of predictability' emerges and reliable prediction can resume.