Foundations of Technical Analysis. Discussion. The Journal of Finance

main motivations behind the algorithm developed here for pattern recognition is that the techniques currently used by chartists are extremely subjective. What may appear to be a head-and-shoulders pattern to one chartist, for example, may appear to be noise or an altogether different pattern to another chartist. Although the techniques in LMW are closer to objective pattern recognition, they also contain several layers of subjectivity.

Discussion

First, the definitions of various patterns in terms of local extrema and their relative positions are subjective. Second, the choice of a particular smoothing technique is also subjective. Section I.A of LMW lists a variety of smoothing estimators that are available in the literature, and for this paper, the authors have chosen the kernel regression estimator. Finally, the choice of bandwidth parameter h used in the kernel regression is also subjective.

To an extent, subjective choices of this nature are the prerogatives of the researchers and cannot be entirely avoided. The main issue, however, is whether the results are robust with respect to different algorithms. A priori, I expected that robustness would not be a serious issue here, but the results in the paper suggest otherwise. The paper reports that the bandwidths for the kernel regressions obtained by minimizing the cross-validation functions yielded fitted values that the authors deemed to be too smooth based on their discussions with practicing chartists. So the paper uses bandwidths that are 30 percent of the optimal ~in a statistical sense! bandwidths. A potentially interesting avenue for future research would be to examine the extent to which the results are sensitive to various choices discussed in the last paragraph.

A more basic question here is whether in the first place it is necessary to fit a smoothed function to objectively identify patterns. Naturally, there will be many more local extrema in the raw data than in the smoothed data. However, it is possible that some of these extrema can be screened out by appropriate filters ~such as discarding local maxima that are less than x percent above the nearest local maximum!. This is another issue that can be addressed in future work to convince the readers that the price data need to be smoothed to identify patterns on the computer.

III. Evaluating the Informativeness of Patterns

LMW evaluate the information content of the technical trading rules by comparing the return distribution conditional on the occurrence of a particular pattern with the corresponding unconditional return distribution. The extant literature typically judges the usefulness of trading rules by evaluating their profitability because this is the metric that is of most direct interest to investors. Although the paper does not discuss the profitability of these trading rules, the empirical results in the paper allow us to address this issue.

Table I presents the standardized mean returns ~the difference between conditional and unconditional mean returns divided by the unconditional standard deviation! conditional on the occurrence of various patterns for the NYSE0AMEX sample and the Nasdaq sample. The standardized mean returns are remarkably close to zero, and the largest magnitude of standardized mean is only 20.062 ~for BBTOP in the NYSE0AMEX sample!. Clearly, these results do not suggest that any of these trading rules yields profits

Table I

This table presents test statistics under the hypothesis that the conditional one-day mean standardized returns, three days following the conclusion of an occurrence of one of the 10 technical indicators in LMW, equal zero. Standardized return is computed as the difference between the raw return and the unconditional mean return divided by unconditional standard deviation. The data are taken from the tables in LMW indicated in parentheses. See Table I in LMW for descriptions of the technical indicators.

	HS	IHS	BTOP               BBTOP              TTOP              TBTOP					RTOP	RBTOP	DTOP	DBTOP
			NYSE0AMEX Stocks ~1962–1996!
Standardized Mean ~Table III!	20.038	0.040	20.005               20.062			0.021	20.009	0.009	0.014	0.017	20.001
Occurrences ~Table I!	1,611	1,654	725                        748			1,294	1,193	1,482	1,616	2,076	2,075
t-statistics	21.525	1.627	20.135               21.696			0.755	20.311	0.346	0.563	0.775	20.046
			Nasdaq Stocks ~1962–1996!
Standardized Mean ~Table IV!	20.016	0.042	20.009	0.009	20.020		0.017	20.052	0.043	0.003	20.035
Occurrences ~Table II!	919	817	414	508	850		789	1,134	1,320	1,208	1,147
t-statistics	20.485	1.200	20.183	0.203	20.583		0.478	1.751	1.562	0.104	21.185

Discussion

that are of any significance from an economic standpoint. The t-statistics in the table indicate that none of the conditional mean returns are reliably different from zero.

These results indicate that the technical trading rules cannot