Candle Pattern Forecast

by Kyra Packens, 3561 days ago

Share |

This indicator makes a prediction for the returns of the current bar based upon averaging the returns of the NumClosest (parameter) candlestick patterns that were most similar to the current bar's pattern in the Lookback (parameter) period. "Similarity" here is the squared euclidean distance between vectors describing the two patterns.

The Filter (parameter) can be useful for training the forecast under differing market regimes; any pattern where the filter is false on the final bar of the pattern will be excluded from the "training" set. The indicator will also produce a prediction of 0 for any bars where the filter is false.

Examples:

kpCandlePatternForecast(1000, 3, 50, 1) -> would form a prediction by averaging the next day returns of the 50 3 bar candle patterns most similar to the current 3 bar candle pattern found in the previous 1000 bars.

kpCandlePatternForecast(1000, 3, 50, close > sma(close, 20)) -> same as above but would only include patterns in the training set if the final bar of the pattern closes above the 20 period SMA. The forecast would also be zero for any bar whose close is below the 20 period SMA.

kpCandlePatternForecast(90, 2, 10, 1) -> would form a prediction by averaging the next day returns of the 10 2 bar candle patterns most similar to the current 2 bar candle pattern found in the previous 90 bars.

Interpreting the results:

Any bar where the indicator produces an excessively high or low result may be a good signal to go long / short for the next day as similar patterns have generated an unusually high / low return over the lookback period. Playing with the filter parameter can also be a good way to visually get an idea of the tendencies of an instrument for the filter regime over the lookback period.

Important: Forecasted returns are from close of the current day to the next day's close. Make sure to SetSimTiming appropriately when you're backtesting.

Credit:

K Nearest Neighbors is certainly not a novel idea, but this page:

http://qusma.com/2013/04/29/mining-for-three-day-candlestick-patterns/

Got me interested enough to code up the indicator for QS. Have a look there for a nice writeup including performance statistics as well as some ideas for improving things even further.

Happy Trading!
-- kp