ACPR: a breakthrough chart pattern recognition tool

This article will introduce what we are calling automated chart pattern recognition or ACPR. It is something we are very excited about because it is using cutting edge innovation to significantly improve our trading here at GDA. In order to understand how ACPR will help us, however, it is first necessary to explain how we trade.

Our Current Approach to Trading

At GDA, we have adopted a systematic discretionary trading methodology. This approach involves a process that a trader systematically completes when trading. There are three steps within this process:

1. Market analysis
2. Trade plan design
3. Execution

Why technical chart pattern analysis does NOT work

It is important to understand that crypto is not like forex or equities and technical chart pattern analysis does not work like it does in those markets. This is so for a few reasons. Firstly, crypto markets have an unusually high reliance on retail traders who trade differently compared to other markets. Oftentimes it is because of a lot of emotional investment going on. Furthermore, crypto has what can be called a tribal phenomena — participants operate in groups much of the time, whether it be trading, crowdfunding or something else. Most importantly however is because the crypto space remains unregulated. This allows large market participants to manipulate the market and abuse smaller traders. Some tools that are designed to identify these chart patterns are developed and released by these big players for the sole purpose of misleading traders. If you have enough power or influence to make a portion of the participants act the same, then you can exploit that.

Our market analysis

While these “typical technical chart patterns” are still relevant, we understand their problems which is why it is only one piece of our market analysis puzzle. We also recognise an asymmetrical relationship between liquidity and market sentiment and direction. When liquidity is high, market participants are feeling confident and entering more trades and larger trades. That is why we afford much of our attention to market liquidity and volume — and whether liquidity is positive or negative. We use proprietary filters and formulas to reduce the false noise in the data and provide our traders with a clearer picture of the market (for more information on this, check out our other article). Our data processing techniques have been tried and tested by our own team of traders for years and we know they are very accurate.

Using this advanced data, our traders are better equipped with superior market understanding. A trader’s market analysis involves two components, typical technical chart pattern analysis and liquidity and volume cluster analysis.

figure 1: Technical chart patterns

figure 2: Liquidity/volume cluster chart

This analysis helps traders identify correlations and necessary market levels which help them understand what the best trade will be. Additionally, traders will apply proprietary patterns in their market analysis generated by the two components discussed above.

figure 3: Proprietary pattern example

These three components — typical technical chart patterns, volume and liquidity clusters, and proprietary patterns — make a trader’s market analysis.

Trade plan design

Following a trader’s research and analysis, they then create a trade plan. This plan sets out all the necessary factors of a trade.

figure 4: An empty trade plan

The trader will determine trade construction parameters — such as entry prices, stop losses and take profits, equity allocation and position hold duration — finance parameters — such as risk, PnL and Sharpe ratio — and performance parameters. Based on the trader’s analysis, they can produce multiple trade plans, with each trade compensating the risk of other trades.

Trade execution

This step is relatively simple. Based on the trade plans developed by the trader, they then go and execute according to the set parameters. The systematic approach throughout the first two steps helps make the trade execution very simple and stress-free.

We are not claiming to have invented this approach to trading, but it works very well. Our proprietary insights into the market help improve this trading approach to make it even more profitable.

ACPR: A breakthrough ML model

This trading approach, while we have found it to be quite successful, is hindered by the amount of manual analysis and human intuition. The superior data from our proprietary market insights is a valuable tool the traders use, but it can be made even better through an implementation of machine learning to replace traders’ market analysis. We not only envision that this model will be able to more accurately identify market parameters, but we also see it being an innovative tool across the industry. This extraction of information from charts is a perfect opportunity to innovate with machine learning. The model would be able to do much of the ground work involved in the market analysis and trade plan design stages and its output would automatically identify levels in the market structure and assign values in the trade plans.

The concept of a chart pattern recognition tool is not new. Such tools exist currently that can be used to identify chart patterns in the market, however they are severely limited by the data they use. Because raw market data can be so unreliable, these tools that analyse them are equally unreliable. We recognise this shortfall which is why ACPR won’t just be like these other chart pattern tools. GDA’s superior proprietary data will be the basis of ACPR, overcoming this problem. A tool like ACPR does not currently exist and we are so excited to be working on this breakthrough project.

Model design

Innovations and development of machine learning over the last decade has been enormous. One notable subset of machine learning that has improved significantly is image classification and the use of convolutional neural networks (CNN). That is why ACPR will be a CNN with LSTM cells, shown to be a great architecture for pattern recognition in images.

figure 5: A CNN

Although we refer to it as a single model, it will be implemented using three individual models; one for each of these image classes:

• Typical technical chart patterns (see figure 1)
• Crypto-native liquidity and volume clusters (see figure 2)
• Proprietary configured GDA technical patterns (see figure 3)

This three part design is attempting to mimic the process traders follow to create trade plans.

Data

The first step of developing this model is building a training dataset. Given the unique nature of this project, no labelled datasets currently exist. Our own data set is being created, consisting of high-resolution images labelled with patterns and levels.

Use Cases

Of course there is the obvious use case that we propose to use this model for primarily and that is within our trading approach. There are other use cases however which we will be pursuing. There is no reason for us to develop this tool and then keep it behind closed doors. This kind of innovation is exciting to us and we want to share it with everyone. We want to offer this model so that others can use it for their own market analysis. Its possibilities aren’t limited to what we use it for in our own trading strategies. New patterns, levels or other market features can be identified by training the model on new datasets. The model can be directly implemented into a trading bot as well and used as a signal generator or connected to an alert channel on telegram for example to receive notifications.

There are many other avenues of innovation that are possible and we are so excited to see what it will be capable of. Please stay tuned for future updates and follow us if you don’t want to miss an article.