5.1. Conclusion
This study builds three simple order-driven zero intelligence agent-based artificial stock market models. These models are tailored to the trading rules employed by TWSE.
The samples of this study cover different investment vehicles and various characteristics. They are as follows: Taiwan Top50 Tracker Fund (0050.TW), Polaris/P-shares Taiwan Dividend+ ETF (0056.TW), Cathay No.2 Real Estate Investment Trust (01007T.TW), Gallop No.1 Real Estate Investment Trust Fund (01008T.TW), China Steel (2002.TW), TSMC (2330.TW), MediaTek (2454.TW), HTC (2498.TW), President Chain Store (2912.TW), Inotera (3474.TW). We use data from February 2008 to May 2008.
We compare the simulation transaction cost with actual transaction cost. All actual transaction costs are smaller than 3%, but the simulated transaction costs are much higher. The results show that investors do not submit market orders blindly, they may split a market order with large order size into many market orders with small order size or they may wait for the appearance of the opposing orders. But if an investor is eager to liquidate his holdings, i.e. he does not want to split his orders or wait for a long time, our models would be useful tools for transaction cost estimation in this situation. We also find that the DFGIS model performs well in frequently traded securities and DFGIS-ACD model performs well in securities not traded frequently.
The simulation results shown in Table 10 and Table 11 can be reference for the investors of these 10 securities. For example, if an investor’s shareholding is 15% of daily trade volume, he can liquidate his shareholding in one trading day without any transaction cost on the average. If he holds more than 50% of the daily trade volume and want to liquidate in a trading day, the maximum transaction cost may be 7%.
Limitation of this study is as follows: the models in this study can only estimate the transaction cost in common market condition. Joulin, Lefevre, Grunberg, and Bouchaud（2008）analyzed 893 stocks in NASDAQ and NYSE and found the relation between price jumps and news in Dow Jones using data from 2004 to 2006. They found that the price jumps are much more than the news. They believed that the spontaneous price jumps come from vanishing liquidity. The explanation of vanishing liquidity is as follows: liquidity providers place their limit order cautiously, and tend to cancel orders when uncertainty signals appear. Our models cannot estimate the
25
transaction cost in extremely pessimistic market condition with vanishing liquidity.
5.2. Suggestions for further research
Firstly, the parameter estimation in the DFGIS-ACD model using EACD(1,1) is a tentative experiment, it does not mean that EACD(1,1) is the best model. Although there are many extensions of ACD model, very few studies compare different ACD models using the same data (Pacurar, 2008). Hence, in the future we can incorporate different extensions of ACD model into our models.
Secondly, we assume the order size in DFGIS model and DFGIS-ACD model follows half-normal distribution, and in DFGIS-ACD-ACD model follows
exponential distribution. We find that the frequency of large order size in half-normal distribution is obviously smaller than actual data. As Fig. 5 shows, in the
DFGIS-ACD model, the order sizes are less than 300,000 shares, but the maximum order size could be 499,000 shares in reality. The growth of algorithmic trading in recent years results in the high frequency of specific order size (for example, 10, 20 or 50). For example, Fig. 6 shows that some market participants of Taiwan Top50
Tracker Fund (0050.TW) split their orders into many small orders with order size 10.
In the future, we can simulate the order size using historical distribution instead of a common probability distribution.
Fig. 5 Comparison of order size between simulation and historical data. (Taiwan Top50 Tracker Fund (0050.TW) on March 20th 2008, and we only show the part in which order size larger than 50,000 shares)
Fig. 6 Split large orders into many small orders with order size 10. (Taiwan Top50 Tracker Fund (0050.TW) on March 20th 2008, and we only show the part in which order size smaller than 100,000 shares)
Thirdly, the models of this study have high flexibility, and it is easy to extend the models in many aspects. A more subtle calibration of model parameters may give better results. For example, if we would like to simulate the effect of algorithmic
27
trading, we can calibrate the order size parameter of market order and limit order respectively so that the market order and limit order have their own model parameters respectively in simulations. The parameters of ACD model can be estimated subtly so that the difference between simulated durations and actual durations can be eliminated.
We can also calibrate the model parameters of buy order and sell order respectively so that we can simulate transaction cost in different market conditions, for example, strong buying and strong selling.
Another point is that the result of this study is “a priori knowledge”. We can further get “a posteriori knowledge” by performing experiments on different trading strategies in our models, and analyzing the transaction costs of these trading
strategies.
Finally, in order to achieve further improvement on efficiency of trading, TWSE changes from call auction to continuous auction step by step. The matching rule of warrants has employed continuous auction since June 28 2010. Therefore, the models of this study should be changed to continuous auction in the future. Furthermore, comparison of transaction cost between call auction and continuous auction can be made in order to be the reference for stock exchange when designing matching rules.
References
[1] Introduce of trading rules in the stock exchange market, Taiwan Stock Exchange Corporation, http://www.twse.com.tw/ch/trading/introduce/introduce.php。
[2] Almgren, R., and Chriss, N., 2000, Optimal execution of portfolio transactions, Journal of Risk, 3：5-39.
[3] Basel Committee on Banking Supervision, 2009, International framework for liquidity risk measurement, standards and monitoring, Consultative Document, Issued for comment by 16 April 2010, Bank For International Settlements.
[4] Daniels, M. G., Farmer, J. D., Gillemot, L., Iori, G., and Smith, E., 2003.
Quantitative model of price diffusion and market friction based on trading as a mechanics random process. Physical Review Letters, 90
（10）:108102-1-108102-4.
[5] Engle R. F. and Russell, J. R., 1998. Autoregressive conditional duration: A new model for irregularly spaced transaction data. Econometrica, 66(5):1127-1162.
[6] Farmer, J. D., Patelli, P., and Zovko, I. I., 2005a. The predictive power of zero intelligence in financial markets. PNAS USA, 102(11): 2254-2259.
[7] Farmer, J. D., Patelli, P., and Zovko, I. I., 2005b. Supplementary material for
‘the predictive power of zero intelligence in financial markets’. PNAS USA, 102(6).
[8] Guo, T. X., 2005. An agent-based simulation of double-auction markets, Master Thesis of Graduate Department of Computer Science, University of Toronto.
[9] Hetland, M. L., 2008. Beginning Python: From Novice to Professional（2nd ed.）（pp.153）. Berkeley, CA：Apress.
[10] Joulin, A., Lefevre, A., Grunberg, D., and Bouchaud, J-P., 2008. Stock price jumps: News and volume play a minor role, Working paper,
http://arxiv.org/abs/0803.1769
[11] Manganelli, S., 2005. Duration, volume and volatility impact of trades, Journal of Financial Markets, 8: 377-399.
[12] Meitz, M., and T. Teräsvirta, 2006. Evaluating models of autoregressive
conditional duration, Journal of Business & Economic Statistics, 24: 104-124.
[13] Pacurar, M., 2008. Autoregressive conditional duration (ACD) models in finance: A survey of the theoretical and empirical literature, Journal of Economic Surveys, 22(4):711-751.
[14] Shetty, Y. and Jayaswal, S., 2006. Practical .NET for Financial Markets.
Berkeley, CA：Apress.
[15] Tsay, R. S., 2008. Duration Models, Handout of Business 41910: Time Series Analysis for Forecasting and Model Building at the University of Chicago Booth School of Business.
http://faculty.chicagobooth.edu/ruey.tsay/teaching/uts/lec15-08.pdf
[16] Weisstein, E. W., 2005. "Half-Normal Distribution." From MathWorld--A Wolfram Web Resource.
http://mathworld.wolfram.com/Half-NormalDistribution.html