A reality check on technical trading rule profits in the u,s, futures markets, страница 17

[1] There are two other approaches to handle data snooping problems in studies of technical trading profitability. The first is to replicate models used in a previous study on a new set of data (e.g., Sullivan et al., 1999). However, a replication study is by definition limited to the models and markets analyzed in the original study, and thus may not uncover profitable models in dynamic markets (Timmerman & Granger, 2004). The second is to apply the genetic programming technique by which optimal trading rules are selected ex ante (e.g., Allen & Karjalainen, 1999; Neely, Weller, & Dittmar, 1997). This approach may avoid data snooping biases from using popular trading rules chosen ex post. But, applying a relatively new search technology, such as genetic programming or artificial neural networks, to the sample period before its discovery can be another form of data snooping (Cooper & Gulen, 2006; Timmermann & Granger, 2004).

[2] Excluded nearby contracts in these four markets are September contracts for corn, August and September contracts for Soybeans, and January, February, April, June, August, October, November contracts for copper and silver.

[3] P_t may differ depending on the execution price of a trade, e.g., today’s closing price, tomorrow’s open price, or a daily stop.

[4] Two moving average systems in Lukac et al. (1988), the Simple Moving Average with a Percentage Price Band (MAB) and the Dual Moving Average Crossover (DMC), were integrated into the Moving Average Crossover (MAC) system in this article. Thus, the total number of technical trading systems tested is 14 rather than 15.

[5] Neely et al. (1997) showed that applying higher transaction costs to in-sample periods and lower transaction costs to out-of-sample periods may reduce the problem of over-fitting in-sample associated with high trading frequency.

[6] Figures for the remainder of the markets are available from the authors on request.