Skip to content
All library documents

Evaluating Candlestick Signals Beyond Directional Accuracy

Article Quant Q&A · Author: Aditya Kulkarni
Read the original

Summary

The document discusses adding candlestick-pattern features to a machine-learning model that predicts whether a stock’s next daily close will rise or fall. The researcher scores patterns by the share of observed cases in which price moved in the pattern’s indicated direction, and asks whether patterns near or below 50% accuracy should be discarded.

The response argues that hit rate alone does not determine trading value. Evaluation should include realized risk-reward, average gains and losses, and expectancy per trade, which combines average payoff sizes with the hit rate. A strategy can be profitable with only half its predictions correct if winning trades outweigh losses; a high hit rate can still lose money when losses are much larger than gains. The exchange gives no test results or guidance on validation, transaction costs, or avoiding overfitting, so pattern scores should be treated as only one part of a broader simulated-trading assessment.

Key ideas

  • Directional accuracy alone does not establish whether a candlestick feature is useful for trading.
  • Measure realized average gains and losses and the risk-reward ratio in a simulated strategy.
  • Trade expectancy combines average payoff sizes with the probability of winning or losing.
  • A lower hit rate can be profitable when gains outweigh losses, while a high hit rate can still lose money.
  • The document does not address validation design, trading costs, or overfitting controls.

Tags

EquitiesMachine learningBacktesting
Full text
# Using candlesticks for Stock price direction prediction


# Using candlesticks for Stock price direction prediction












I am working on a college project wherein I want my machine learning model to predict the one-day-ahead direction of a given stock (i.e. whether the closing price of the stock would rise or fall as compared to previous day's closing price).

I am currently working on feature generation/extraction. In stock price direction prediction literature, the use technical indicators has been extensively studied. But I could not find much literature on the use of price action (candlestick patterns, to be specific) for prediction. So I want to implement candlestick patterns along with technical indicators to predict the direction. I generated some candle patterns from my data and assigned them scores.

The scores were assigned as follows:

If r denotes the no. of times the price moved in the direction indicated by the pattern and w denotes the no. of times the price moved in the opposite direction then, score = r/(r + w)

Now this is where I am confused and unsure about my approach. Some patterns obtained scores around 50% or less. Would these really help the model in predicting better? Or should I just drop the idea of using candlesticks and get on with technical indicators only?

Any advice or help is highly appreciated. Thank you.

P.S: I had asked this question on Cross Validated Stack Exchange and was advised to post the question here.

## Answer by amdopt (score 3)

https://quant.stackexchange.com/a/63221

The accuracy of a model is only 1 factor in determining usefulness. Aside from the accuracy, it would help to determine how you would implement it in a simulated trading environment and look into the performance further.

Aside from a hit ratio or accuracy, you should compute other metrics such as:

- The risk-reward ratio that your model realizes (not the theoretical one that you hope it will)



- Average gain and average loss

- Your expectancy per trade: (avg gain x hit ratio) - (avg loss x (1-hit ratio))



There are other metrics, too, but they wind up having some redundancy. I'm sure you could find a ton if you do some searching or googling. The point is that a correct prediction 50% of the time could be profitable if it returns more than it risks. On the other hand, a model that hit's 80% of its trades but only returns 1 for every 10 it risks could be very unprofitable. You should make looking into other metrics a habit to be sure you aren't throwing away models that warrant further investigating or chasing models that aren't worth it.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.