In this paper we have taken the relationship suggested by the technical indicator known as a stochastic to establish cointegration relationships in daily exchange rate data consisting of Close, periodic High and periodic Low prices. Using the dynamic modelling strategy of Clement and Mizon (1991) and Hendry and Mizon (1993) and Johansen (1988) we were able to derive fully dynamic forecasting models for USDDEM and USDJPY, which proved to significantly outperform a random walk, in an out-of sample forecasting experiment, at a time horizon as short as one day, and the models were also demonstrated to have directional forecasting ability.
By transforming our forecasting model into a trading model we were further able to investigate the model’s profitability. The results were compared to a Buy&Hold benchmark, as well as to three different trading strategies commonly used by Technical Analysis. Two of these technical indicators represented variations of the Stochastics and thus allowed us to directly compare the forecasting performance of our model to its generic root. The third trading system was an arbitrarily chosen moving average system, which represents the class of trend-following trading models widely used by technical analysts.
The results of our profitability study showed that while the arbitrarily chosen technical indicators had problems in beating the Buy&Hold strategy, our models had no difficulties in passing this criterion for both currencies. Adjusting for risk, using the methodology proposed by Thomas and Levich (1993), also had no substantial effect on the profitability of our models. Even though the trading strategy of the traditional stochastics and moving average models resulted in a quite high annualised rate of return, they could not match the performance of our models.
The dynamic modelling strategy utilised in this paper must therefore possess an important informational advantage over such models and, indeed, our forecasting analysis revealed that the good forecasting performance of our models is directly linked to the inclusion of error correction components. We of course appreciate that our relatively small sample size means that we have to be extremely careful in interpreting our results.
However, although this might be a serious short-coming for the persistent profitability analysis of the three Technical Trading Models, the high dealing frequency of our models meant that we were able to analyse 250 trading signals, which is in fact similar to the number of trading signals studied in the ‘long-term’ studies of Schulmeister (1987) and Menkhoff (1995). Since the profitability of our models persisted in two completely different market situations, we assign a certain statistical meaningfulness to our results.
Prof. Ronald MacDonald, Prof. Norbert Fiess
Next: Appendix
Summary: Index