Market Timing with Candlestick Technical Analysis¶
Motivation(s)¶
The majority of academic literature have demonstrated that technical analysis centering around trading rules (e.g. moving average, support and resistance, range break-out) does not have value once transaction costs and risk adjustment are taken into account. In contrast, the entire investment industry have adopted candlestick technical analysis for short-term horizons.
Proposed Solution(s)¶
The authors analyze the profitability of candlestick technical analysis in order to reconcile the disparity between academia and industry.
Evaluation(s)¶
The authors examine holding periods of two, five, and ten days using stocks from the DJIA during the 1992 - 2002 period. The price data came from Reuters and dividend data were sourced from CRSP. Note that the different holding periods produced similar results.
The analysis focused on \(t\)-statistics: compare the returns following a technical analysis signal to returns when there is no signal. Returns were measured on a daily basis as the log difference of price relatives. To handle skewness and leptokurtosis, they applied the bootstrapping methodology and fitted the data to null models such as random walk, AR(1), GARCH-M, and EGARCH. Note that the different models produced similar and consistent results.
The authors used a 10-day EMA to determine bullish and bearish reversal patterns. They executed their trades at the close on the day after a signal; entering at the open price on the day after a signal gave similar results.
The proportion of times that a trading rule produces more profit on the bootstrapped series than on the original series following a signal is a simulated \(p\)-value for the null hypothesis that the trading rule has no value. For instance a bullish candlestick has statistically significant forecasting power at the 1% level if the simulated \(p\)-value is less than 0.01. Put another way, more profit should be produced on the random series than the original less than 1% of the time. For a bearish candlestick to have forecasting power at the 1% level the simulated \(p\)-value should be more than 0.99. In other words, there should be more profit on the random series than on the original more than 99% of the time.
The results show that candlestick patterns do not yield statistically significant returns except for the Opening White Marubozu, Long Black, and Black Marubozu. The returns following Opening White Marubozu are negative, which is the opposite to what candlestick technical analysis theory suggests. Likewise, contrary to candlestick theory, the bearish Long Black and Black Marubozu indicate higher than average returns over the next ten days. Note that none of the bootstrap results are statistically significant; this indicates that the \(t\)-statistic’s assumptions is being violated, and thus the foregoing three patterns may be a coincidence.
Future Direction(s)¶
Technical analysis is widely used by humans and possibly algorithms. How to incorporate this self-fulfilling prophecy into a neural network or contextual bandit strategy?
Question(s)¶
The authors mentioned that it is not appropriate to consider the daily returns of candlestick technical analysis on an annual basis. Why wouldn’t backtesting with this strategy be a reasonable binary indicator of profits?
Analysis¶
Candlestick lines and patterns are supposed to predict larger than normal positive returns, but sometimes actually predict smaller than normal returns; the converse also holds. However, there are only weak evidence suggesting candlestick lines and patterns generate predictability in prices.
The analysis is centered around statistical significance of each candlestick pattern and completely ignores the possible use of each pattern as an indicator. The paper would be more interesting if the authors examined how these weak indicators could be combined to predict trends instead of prices (e.g. contrarian and momentum strategies).
One interesting point is that three of the candlestick patterns contradicts the original candlestick theory, which seems to hint that technical analysis should be more data-driven. The summary statistics demonstrate that instead of trying out different statistics, just focus on profit margins.
References
- MYR07
Ben R Marshall, Martin R Young, and Lawrence C Rose. Market timing with candlestick technical analysis. Banking and Finance, 2007.