Fortunately, statistical methods combined with computer power can be a good solution to make the candlestick patterns recognition works less time-consuming and more effective. These days, Neural Network (NN) can help to automate a candlestick patterns recognition task. NN should be properly trained in order to be able to predict the next candlestick parameters with the highest probability. One of the obvious problems of implementing a candlestick pattern NN predicting system is a formalization of inputs.

Numerous tests show many possibilities of improving NN candlestick patterns recognition systems. For example, output result can be composed from selected optimized calculations based on different historical periods. Also there are many different ways to formalize the shapes and relative positions of candlesticks. As well as, as back- and forward-testing show, it is reasonable to re-train NN for each particular type of shares and latest historical periods to make the forecast more accurate.

Optimal Solution. There is an automated tool FTA-2 (free use of fully-functional version for one month). It has module which enables using Neural Network to recognize typical candlestick patterns and predict future prices (open, high, low, close). This module predicts only one next candlestick but the prediction can be successfully used for different widths of candlestick, i.e., the number of trading days in one candlestick. The module has been enhanced to calculate result that is composed from different historical periods that allows making the forecast more accurate. Also it can perform comparative forecast analysis for many symbols.

Useful resources:

• Candlestick basics – major signals
• Neural Network basics – introduction
• The computer program which enables using Neural Network to recognize typical candles patterns and predict future prices – software FTA-2

SAR indicator gives a strong signal when a price trend is about to reverse, therefore, this indicator can be used for prediction. To compare the predictive ability of SAR with other indicators, it has been implemented into the technical analysis module of Fundamental-Technical Analyzer FTA-2. SAR calculations have been used to collect statistics based on the forecast simulations for major indexes and ETFs during August-October 2011 period. As a result, SAR’s position was mostly in “top ten” indicators list.

The research and presented chart are made by Fundamental-Technical Analyzer FTA-2, one of the software modules that enables composing Neural Network forecasts of many indicators with weights accordingly to each indicator’s predictive ability.

Omitting logical rules for accelerating factor and reversal conditions, a recurring core formula for SAR is the following:

SAR (current point) = AF * [EP – SAR (previous point)] + SAR (previous point)

where:
AF – Acceleration Factor (normally starts from 0.02 and increases by 0.02 if each next point reaches a new extreme, saturates until 0.2);
EP – Extreme Point (lowest low or highest high).

To summarize, Parabolic SAR can be enriched by combining it with Neural Network and successfully used for predicting stock market prices. Combing it with Neural Network allows extracting more statistically stable patterns and, therefore, providing a better accuracy in the forecast. As simulations showed, improved results can be achieved if SAR is transformed into more sensitive indicator by subtracting it from close price (it indicates the degree of SAR and price convergence).

Among studied indicators are Relative Strength Index (RSI) and Moving Average Convergence/Divergence (MACD) indicators. They have been transformed to a slope of line and differentially-coupled with a price line slope. Indicators and price transformation to line slopes has been performed using Least Squares Linear Regression within a sliding 10-day period (moving window). A comparative analysis of indicators’ predictive abilities showed that these two coupled ones are better than around 90% of all other (57) the most popular indicators.

The chart below shows an example of such forecast:

The research results presented here have been calculated using a feature of Technical Analyzer TA-1 (TA). Except analyzing chart with indicators and historical data, TA enables to perform a 10-day forecast using Artificial Neural Network. The calculations can be done on the basis of one selected technical indicator or all available. If all indicators used, TA decides how much weight should be assigned for each indicator’s forecast in a composite result by using back-testing for particular market conditions and a specific stock. Each weight is proportional to a predictive ability of a corresponding technical indicator.

The charts below shows how actual 5-day performance (yellow line) differ from predicted performances by these two methods. The top half is the comparison of Neural Network prediction, bottom half – Cycle Analysis. Green bars mean buy signals, red – sell*.

Three major conclusions for this particular historical period:

1. Cycle Analysis prediction gives signals too early, Neural Network prediction – too late.
2. In average, the prediction by Cycle Analysis showed slightly better accuracy than the one by Neural Network.
3. It seems logical to combine these two methods to improve the accuracy.

* ) The calculations have been performed by an integrated experimental system that combined two applications: Cycle Analysis predictor SMAP-3 and Neural Network predictor NNSTP-2