Return predictability and the ‘wisdom of crowds’: Genetic Programming trading algorithms, the Marginal Trader Hypothesis and the Hayek Hypothesis

https://doi.org/10.1016/j.intfin.2015.02.009Get rights and content

Highlights

  • New evidence of return predictability.

  • Comparison between STGP and benchmarking models.

  • Short-term outperformance of the STGP.

  • Evidence supporting the Hayek Hypothesis.

  • We find little support for the Marginal Trader Hypothesis.

Abstract

We develop profitable stock market forecasts for a number of financial instruments and portfolios using a special adaptive form of the Strongly Typed Genetic Programming (STGP)-based trading algorithm. The STGP-based trading algorithm produces one-day-ahead return forecasts for groups of artificial traders with different levels of intelligence and different group sizes. The performance of the algorithm is compared with a number of benchmark forecasts and these comparisons clearly demonstrate the short-term superiority of the STGP-based method in many circumstances. Subsequently we provide detailed analysis of the impact of trader cognitive abilities and trader numbers on the accuracy of forecasting rules which allows us to conduct new experimental tests of the Marginal Trader and the Hayek Hypotheses. We find little support for the Marginal Trader Hypothesis but some evidence for the Hayek Hypothesis.

Introduction

A voluminous literature exists on predicting stock market returns using different variables. However, most of the empirical studies suffer a major shortcoming – they find evidence in favour of in-sample return predictability but no presence of out-of-sample profitability. Goyal and Welch (2008) suggest that the predictive ability of nearly all stock return forecasting models does not hold up in out-of-sample validation. Moreover, a number of researchers argue that many studies finding predictability of stock returns might be spurious. This is based on the empirical assumption that the persistence of the forecasting variables and the correlation of their innovations with returns might bias the regression coefficients and negatively affect t-statistics (Stambaugh, 1999, Lewellen, 2004). These issues are exacerbated when researchers are data mining, taking into account many variables, and reporting only statistically significant findings (Ferson et al., 2003, Ferreira and Santa-Clara, 2011). Most prior studies investigate the relationship between the available information and the stock returns using simple linear regression. This is quite a restrictive approach as there is no evidence to support the assumption that the relationship between stock returns and the financial and economic variables in these studies is perfectly linear (Enke and Thawornwong, 2005).

In this paper we use Strongly Typed Genetic Programming (STGP), a new forecasting approach to predict stock market returns, based on intelligent agents interacting in an artificial stock market. Research on financial markets has rarely used STGP models although they offer considerable advantages compared to any other forecasting techniques including the ability to evolve through time and deal with non-linear relationships (see Manahov et al., 2014a, Manahov et al., 2014b, for recent papers describing STGP models and applying them in financial markets). We compare the STGP predictions to several appropriate benchmark forecasts on an in-sample and out-of-sample basis using a number of performance measures to investigate whether, to what extent and in which form, stock returns in excess of the risk free rate are indeed predictable and profitable compared to a buy-and-hold strategy with dividends and without dividends. This is done by measuring and quantifying the exact level of generated profit, taking into account the transaction costs of 85 basis points per round trip.

The other important question under investigation in this study relates to how stock markets incorporate information and generate prices. There are two competing hypotheses about how markets work – the Marginal Trader Hypothesis (MTH) and the Hayek Hypothesis. While the MTH postulates that a small fraction of savvy individuals are capable of setting market prices, the Hayek Hypothesis suggests that markets can operate in an efficient manner even when market participants have a limited knowledge of the surrounding environment. Blackwell (2011) and Blackwell and Pickford (2011) argue that the MTH is so intuitive to economists that it seems never to have been tested directly and examined rigorously. In real financial markets it is difficult to determine whether the market price is actually the ‘correct’ price for any asset because is impossible to know the ‘correct’ price (Blackwell, 2011). Artificial stock markets do not necessarily suffer from this limitation because the market price of the instrument in the market can reflect some underlying prediction and the accuracy of that prediction can be checked. In addition, in real markets it is clearly not possible to directly control the number and intelligence of the participants. Using our STGP models we conduct experiments which avoid the difficulties in testing the MTH and the Hayek Hypothesis and provide appropriate laboratory conditions for direct examination of both hypotheses. To test the MTH we investigate whether a small fraction of the genetic population, the ‘Best Performing Traders’, are capable of generating more accurate forecasts. We show that the forecasting process emerging from the collective behaviour (competition and co-evolution) of the entire set of traders in the market is a better predictor than that produced by any small subset of traders, even the ‘Best Performing Traders’. We demonstrate that enhanced genetic diversity leads to increased accuracy in forecasts because of more heterogeneous trading rules and behaviour. To examine the validity of the Hayek Hypothesis we populate another model with ‘Low Intelligence Traders’ in order to investigate their forecasting abilities. We provide evidence that the level of traders’ intelligence does not necessarily play a significant role in improving the quality of forecasts (this represents new evidence favourable to the Hayek Hypothesis in a broader environment). The reminder of this paper is organised as follows: Section 2 presents the background to the study and reviews the literature in the field; Section 3 discusses the experimental design; Sections 4 Simulation results – relevant for market prediction, 5 Simulation results – relevant to assessing the ‘wisdom of crowds’ discuss the simulation results relevant for prediction and assessing the ‘wisdom of crowds’ respectively and the paper concludes with overall remarks and discusses avenues for further research in Section 6.

Section snippets

Background

In this section we outline the background to the study and the relevant prior literature. In Section 2.1 we discuss the forecasting of financial markets and how Genetic Programming has previously been used in this domain. Section 2.2 outlines prior research on the implications of trader cognitive abilities for forecasting accuracy.

Developing initial trading rules

We use a STGP based artificial stock market model similar to that used in Manahov et al., 2014a, Manahov et al., 2014b and the most salient features of the model are described in those papers. For conciseness we refer readers to the previous papers for details and explanation of the model and here confine ourselves to describing the particular features of the model used in the experiment reported in this paper. Table 1 shows the main parameters of the STGP model used here.

Artificial stock market structure

We study traders’

Simulation results – relevant for market prediction

This section deals with the elements of our results that are particularly related to predictive accuracy. Section 4.1 outlines the results from our main experiment and Section 4.2 gives some robustness results on some constructed portfolios.

Simulation results – relevant to assessing the ‘wisdom of crowds’

This section deals with the elements of our results that are relevant for assessing the ‘wisdom of crowds’. Section 5.1 outlines the evidence about the MTH and Section 5.2 gives evidence about the Hayek Hypothesis.

Conclusions

Developing and selecting a plausible model are the two most important ingredients of the forecasting process. In this paper we compare the forecasting and in-sample performance of an STGP model with that of several benchmarking models. We apply formal statistical model selection criteria to establish the most accurate model for modelling the returns of a number of financial instruments and portfolios. We place heavy emphasis on genuine out-of-sample (ex ante) predictions. Our findings

References (54)

  • R. Stambaugh

    Predictive regressions: theoretical results and applications

    J. Financ. Econ.

    (1999)
  • K. Beckmann et al.

    Markets and the use of knowledge.

    Testing the ‘Hayek hypothesis’ in experimental stock markets.

    (1994)
  • L.A. Becker et al.

    GP-evolved technical trading rules can outperform buy and hold

  • C. Blackwell

    A quasi-experimental test of the Marginal Trader Hypothesis

    Kyklos

    (2011)
  • C. Blackwell et al.

    The wisdom of the few or the wisdom of the many? An indirect test of the Marginal Trader Hypothesis

    J. Econ. Financ.

    (2011)
  • G. Brüggelambert

    Information and efficiency in political stock markets: using computerised markets to predict election results

    Appl. Econ.

    (2004)
  • C. Camerer et al.

    The curse of knowledge in economic settings: an experimental analysis

    Journal of Political Economy

    (1989)
  • M.M. Carhart

    On the persistence in mutual fund performance

    J. Financ.

    (1997)
  • W.-S. Chan et al.

    Robustness of alternative non-linearity tests for SETAR models

    J. Forecast.

    (2004)
  • C. Chatfield

    Model uncertainty and forecast accuracy

    J. Forecast.

    (1996)
  • S.H. Chen et al.

    Option pricing with genetic programming

  • W. Chen et al.

    Decimalization, ETFs and futures pricing efficiency

    J. Futures Mark.

    (2009)
  • N.K. Chidambaran et al.

    Adaptive evolutionary approach to option pricing via genetic programming

  • M.P. Clements et al.

    A Monte Carlo study of the forecasting performance of empirical SETAR models. Working paper

    (1997)
  • D.D. Davis et al.

    The Hayek hypothesis in experimental auctions: institutional effects and market power

    Econ. Inq.

    (1991)
  • E.F. Fama et al.

    Common risk factors in the returns on stocks and bonds

    J. Financ. Econ.

    (1993)
  • E.F. Fama et al.

    Multifactor explanations of asset pricing anomalies

    J. Financ.

    (1996)
  • Cited by (14)

    • Trading support system for portfolio construction using wisdom of artificial crowds and evolutionary computation

      2021, Expert Systems with Applications
      Citation Excerpt :

      Another difference is the framework itself: Manahov proposes a virtual market in which VE trade among themselves; here we propose a virtual community of investors who make a recommendation for a certain period in a real market. Some studies use VE to forecast the future price of a particular stock or index (Alqahtani & Ertekin, 2017; Maknickas & Maknickiene, 2019; Manahov, Hudson, & Hoque, 2015) and they make recommendations regarding the crowd’s stock picks. In this study, the algorithm generates signals instead of potential prices, which provides a more clear evaluation from the investor’s point of view.

    • Grammatical Evolution-based ensembles for algorithmic trading

      2019, Applied Soft Computing Journal
      Citation Excerpt :

      Since Allen and Karjalainen [1] published their seminal piece on evolution of trading rules using Genetic Programming (GP), many authors have made related contributions either based on the same technique, or Grammatical Evolution (GE). Among them, [2–9]. Most of these contributions generate investment rules based on a combination of raw market data and technical indicators and, unlike related approaches that use genetic algorithms or evolution strategies to optimize predefined rules, these have the advantage of creating flexible structures automatically.

    • Evolution of trading strategies with flexible structures: A configuration comparison

      2019, Neurocomputing
      Citation Excerpt :

      This approach imposes data type constraints and uses generic functions and data types, which overcomes some issues related to the closure requirements of the canonical version of GP. Manahov et al. [33] developed profitable stock market forecasts for a number of financial instruments and portfolios using a special adaptive form of STGP. The STGP-based trading algorithm produced one-day-ahead return forecasts for groups of artificial traders.

    • An in-depth investigation of genetic programming and nine other machine learning algorithms in a financial forecasting problem

      2022, 2022 IEEE Congress on Evolutionary Computation, CEC 2022 - Conference Proceedings
    • Trader-company method: A metaheuristics for interpretable stock price prediction

      2021, Proceedings of the International Joint Conference on Autonomous Agents and Multiagent Systems, AAMAS
    View all citing articles on Scopus
    1

    Tel.: +44 1482463189.

    2

    Tel.: +44 190432783.

    View full text