RainBoltz
commented
3 years ago Hi @shreyas0906 , its glad to see someone is paying attention on our work (though I'm not on the author list because of some reasons lol).
I can conclude the main ideas of this paper for you:
Goal
To see if CNN is capable of learning the implicit features of "Price Trend" from "Candlestick Patterns".
Main Approach
Model
A reduction version of LeNet-5. We don't need a heavy model since our input image is really small.
Data
EUR/USD 1-minute price data. We compared the performances between OHLC and CULR (both of them come from price data in candlestick form, mentioned in section 3.4)
Data Preprocess
We transform price data (candlestick form) into GAF (GASF, more precisely).
Note that OHLC or CULR will construct 4 independent timeseries, so the input will be N*N*L*4 images (N for window size, L for the data length after performing sliding window, not the original data length).
Label
We labeled each individual pattern as a class, so there are 8 classes and an additional class that stands for "Others", total 9 classes.
For further explanation, "Price Pattern Indicator" is an old-fashioned indicator used to predict whether the price trend is going to reverse or going to be stronger. It is presented by trading textbooks.
Here we picked 8 classic patterns (indicators), which are all used to predict the "Trend Reversal" (you can find the illustrations on google):
- ↑
Uptrend Patterns: Morning Star, Hammer, Bullish Engulfing, Inverted Hammer - ↓
Downtrend Patterns: Shooting Star, Evening Star, Bearish Engulfing, Hanging Man
(and so "Others" class indicates "Neutral Trend".)
Feel free to have further discussion or ask me anything! You can also contact the first author @bigmoumou with direct message if you want.
shreyas0906
Hey RainBoltz,
I was reading your paper . Can you please elaborate on how you generated labeled data. I couldn't clearly understand that part in your paper.
Also it would really help if you could share the script you used to generate the labelled data.
Thanks in advance!