Maveryx Success Stories
Learn how Alteryx customers transform their organizations using data and analytics.STORIES WANTED
Showcase your achievements in the Maveryx Community by submitting a Success Story now!
SUBMISSION INSTRUCTIONSPredicting and Trading on the Cryptocurrency Markets using Alteryx
Overview of Use Case
Predict Crypto is a small project, created by a student at the University of Colorado Boulder, that aims to predict, trade and do research on the cryptocurrency markets. The project collects data from a broad range of markets and different data sources, and uses that information to perform predictive modeling and initiate trades. Based on the outcomes of the predictions made, the trading strategy changes to distribute the capital across the assets with the most favorable outlook, and if the outlook for the market as a whole does not look favorable, the funds are invested in the cryptocurrency "USDT", which has a value that always tracks equivalent to the US Dollar, in order to mitigate losses associated with the natural volatility of the cryptocurrency markets.
Describe the business challenge or problem you needed to solve
We were looking to create a fully automated solution with the capability of dynamically pulling the latest data from a database and produce a brand new set of predictive models on a daily basis, and leverage those models to create new predictions every hour and execute real trades on the live cryptocurrency markets. When it comes to navigating the incredibly volatile world of cryptocurrencies, the most encouraged strategy is to "HODL" (Hold On for Dear Life), or in other words to put the investment away and forget about it for a couple of years. During those years there are sure to be several crazy upward and downward price surges and not all of them are going to be easy to predict, but perhaps there are ways for us to observe patterns within the markets to make medium-term predictions and cash in on those trends better than a simple HODL strategy would perform, so that is the objective of the project.
Describe your working solution
The process in Alteryx from beginning to end is outlined in the graphic below:
1. Extract Data:
- Our database is hosted on Google Cloud Platform and is a MySQL database
- For the predictive modeling, we will need to pull a "train" dataset, which will consist of all the data we collected relating to the cryptocurrency markets in the past, and we will limit the data to be from the past 100 days in order to keep the dataset to a reasonable size:
- When we intend to leverage the models after they are created in the alter steps, we will adjust this query to pull data from the latest hour instead of the latest 100 days, using a field created specifically for this reason called "pkDummy":
2. Data Prep:
- Data is extracted from several different sources within the database using the methodology outlined in step 1, and some data manipulation is performed to join all the data together and prepare the dataset for the predictive modeling step:
- It's not worth taking too deep of a dive into the specifics of the data preparation, but it mainly consists of calculating the target variable, in this case the % change in USD price over the next 6 hour period for individual cryptocurrencies, and joining data across different data sources matching both on when the data was collected and for what asset.
3. Make Predictions:
- Because the trades are executed on multiple exchanges, we create individual models for each exchange. Each exchange has an independent model created using the XGBoost framework programmed in R using the R developer tool. Additional models using the R Predictive tools found within Alteryx are also created:
- The XGBoost model itself is outside the scope of this tutorial, but see the Related Resources section at the end of this article for an excellent practical example of how to create an XGBoost model with automated hyperparameter tuning in R.
4. Execute Trades:
- Using the models created in step 3, predictions on the latest/live data are made. Before actually performing the trades, the predictions get finalized applying an 80% weight to the predictions made by the XGBoost models, and 20% weight to the predictions made by the Alteryx models, adjusting them as necessary based on observed performance. After making the necessary adjustments to the predicted values, the data is passed to a Python tool, which leverages the Shrimpy developer API to execute trades on each of the exchanges individually:
- We use the Shrimpy platform to execute our trades because it can be challenging to connect to the individual exchanges through different APIs, and the Shrimpy.io service allows us to connect to several different exchanges using the same API and methodology. There are also additional benefits to using the service, such as "rebalancing", and allowing other users to automatically follow the same portfolio as the one doing the automated trading. Go to the end of the article to the related resources section for more information about the very awesome Shrimpy.io service/platform.
- Once we repeat the process for several exchanges independently, this is what the workflow that performs trades on the latest data ends up looking like:
5. Compare Model Performance:
- Once the predictions are made, that information is written to a database table as a centralized location to then be able to reference to compare the predicted values vs. actuals:
- In this step we create new fields that summarize the DateTime of when the prediction end is supposed to be. For example, if a prediction was made at 3:00PM on November 23rd 2019, we would be predicting the % price change against the dollar from 3:00PM to 9:00PM.
- After writing the predictions on November 23rd 2019 at 3:00PM (just an example), once 6 hours have passed, we can pickup the data using the Prediction End Date Time field and compare those predictions to the real life results and write the data to a table in the database for further analysis:
- Once that process is complete, the performance vs. actuals are visualized as interactive dashboards for internal analysis using an instance of Metabase running on AWS that works very similarly to how a Tableau Online site would work:
Describe the benefits you have achieved
Alteryx provides the best framework for this project by all accounts, and the process would be incredibly challenging to maintain through R or Python alone. Because of the huge time savings from a development perspective that are achieved through the usage of Alteryx, we are able to create a wide variety of tests and organize them and enable/disable different ones incredibly easily. Another thing that can be challenging using traditional tools is passing data from R to Python, and although the same results can be achieved using R Studio and the Reticulate package, being able to connect the data stream between tools in Alteryx provides a more powerful solution. Having the project built out of Alteryx means there are no limitations to our data capabilities as the project evolves and creates a huge amount of possibilities in terms of what can be achieved in a day compared to what could be done programming in R or Python.
Related Resources
We are working on a web page that will provide a hands-on programming tutorial as well as a more in-depth explanation and will add that here once that has been finalized. The company website for the project is https://www.predictcrypto.com/
Shrimpy:
Shrimpy.io is an awesome platform for cryptocurrency trading, here is an introduction: https://help.shrimpy.io/en/articles/2431063-introduction
Some of our favorites articles from the Shrimpy blog:
- https://blog.shrimpy.io/blog/the-best-threshold-for-cryptocurrency-rebalancing-strategies
- https://blog.shrimpy.io/blog/cryptocurrency-index-funds-step-by-step-guide
XGBoost:
- An XGBoost tutorial with hyperparameter tuning in R we recommend: https://rpubs.com/ippromek/336732
- XGBoost Python documentation: https://xgboost.readthedocs.io/
Metabase:
- Quick overview: https://www.metabase.com/docs/latest/getting-started.html
- Setup a Metabase environment on AWS in less than one hour: https://www.youtube.com/watch?v=wmJ02K8LIFk
Hi @koheny01, thanks for stopping by!
In terms of applying this process to trading on the forex markets, you wouldn't be able to re-purpose anything from the workflows themselves, and the trading step happens through a platform called Shrimpy which only supports cryptocurrency exchanges, but the general process would be the same so I am going to outline that instead:
- First you will need to create or find a dataset with historical data for the currencies you will be trading on. In this step already be thinking about the predictions/trades you want to make; how many minutes/hours/days out into the future should the prediction be for? In terms of what you are predicting, it probably makes the most sense to predict the % change against USD (or any other currency as long as you keep it consistent).
- Now that you have a dataset to work with, you can bring it into Alteryx and do your data prep work. For example, here you would go ahead and calculate/implement the % change vs. USD for the specific currencies for the prediction period. As a last step before creating your model, I would also create a split of the data where you only use 80% of the data to train the model and 20% to test how it's performing. Here is a quick screenshot of one way you could achieve this:
(keep an eye on the fact that I am taking the R side of the join to get everything that was in the original data but not found once the 80% sample is taken to get to the remaining 20% of the data)
- Now you want to create the model you will use to make your predictions. This will vary wildly based on what you are trying to achieve, and there are a ton of great resources in the realm of making predictions, but one extremely powerful framework you could test is XGBoost. It tends to perform quite well on a wide range of problems. One way you could get started, is by taking a more "black-box" version of the model by doing automated hyper-parameter tuning for the model; meaning, the code will produce several versions of the models tweaking hyperparameters automatically, and going with the best version of the model. Be careful of overfitting and try to get a better understanding of what works and what does not over time. Below is the code in R used to make the XGBoost model as described:
library(tidyverse) # data manipulation library(mlr) # ML package (also some data manipulation) library(xgboost) train <- read.Alteryx("#1", mode = "data.frame") test <- read.Alteryx("#2", mode = "data.frame") # Fitting XGBoost trainTask <- makeClassifTask(data = train, target = "Target6hChange_Binary", positive = 1) testTask <- makeClassifTask(data = test, target = "Target6hChange_Binary") set.seed(1) # Create an xgboost learner that is classification based and outputs labels (as opposed to probabilities) xgb_learner <- makeLearner( "classif.xgboost", predict.type = "prob", par.vals = list( objective = "binary:logistic", eval_metric = "auc", nrounds = 300, early_stopping_rounds = 30 ) ) # Create a model xgb_model <- train(xgb_learner, task = trainTask) result <- predict(xgb_model, testTask) head(result$data) # Hyper-parameter Tuning # Full list of parameters and explanations: https://xgboost.readthedocs.io/en/latest/parameter.html # Could also run the following code to get a list with value ranges: getParamSet("classif.xgboost") xgb_params <- makeParamSet( # The number of trees in the model (each one built sequentially) # makeIntegerParam("nrounds", lower = 50, upper = 500), # number of splits in each tree makeIntegerParam("max_depth", lower = 1, upper = 10), # "shrinkage" - prevents overfitting makeNumericParam("eta", lower = .1, upper = .5), # L2 regularization - prevents overfitting makeNumericParam("lambda", lower = -1, upper = 0, trafo = function(x) 10^x), # Type of booster makeDiscreteParam("booster", values = c("gbtree", "gblinear", "dart")), # Additional parameters to test makeNumericParam("min_child_weight", lower = 1L, upper = 10L), makeNumericParam("subsample", lower = 0.5, upper = 1), makeNumericParam("colsample_bytree", lower = 0.5, upper = 1), # Test error type makeDiscreteParam("eval_metric", values = c("auc", "rmse", "logloss")) ) # Set Random or grid control. Grid control does 100% of options and is crazy computationally expensive, # but could set less parameter options and do that instead. Change "Random" to "Grid" and pass no parameters # and should work the same otherwise control <- makeTuneControlRandom(maxit = 30) # Create a description of the resampling plan resample_desc <- makeResampleDesc("CV", iters = 5) # Perform tuning tuned_params <- tuneParams( learner = xgb_learner, task = trainTask, resampling = resample_desc, par.set = xgb_params, control = control ) # Create a new model using tuned hyperparameters xgb_tuned_learner <- setHyperPars( learner = xgb_learner, par.vals = tuned_params$x ) # Re-train parameters using tuned hyperparameters (and full training set) xgb_model <- train(xgb_tuned_learner, trainTask) # Make a new prediction result <- predict(xgb_model, testTask) prediction <- result$data %>% select(pkey = id, Target6hChange_Binary = response) %>% # No sorting has happened, so everything still matches up. mutate(pkey = test$pkey) # Output data to Alteryx write.Alteryx(as.data.frame(prediction), 1) #Save xgb model save.image(file="C:\\Users\\Ricky\\Desktop\\CryptoWork2019\\November\\XGBoostModelsByExchange\\KuCoinXGBoostModel.RData")This code was adapted from this excellent tutorial: https://rpubs.com/ippromek/336732
- Now that you have a saved model, you will need to figure out a way to create a fast enough method to create a new set of data, which will need to match the "train" used in step #3 in terms of data preparation steps, to make predictions on (again, in terms of the % change in price over the prediction time period). This is the most challenging step of the process, as in a lot of cases it's not about making a good prediction but about being the fastest one to take action.
- Once you have figured out how to take care of step #4, you will go ahead and make predictions on the new data using similar code to before (once again in R but XGBoost is available in Python):
library(tidyverse) # data manipulation library(mlr) # ML package (also some data manipulation) library(xgboost) test <- read.Alteryx("#1", mode = "data.frame") # Load XGBoost model which is refreshed once a day load("C:\\Users\\Ricky\\Desktop\\CryptoWork2019\\November\\XGBoostModelsByExchange\\KuCoinXGBoostModel.RData") # Create test task testTask <- makeClassifTask(data = test, target = "Target6hChange_Binary") # Make a new prediction result <- predict(xgb_model, testTask) prediction <- result$data %>% select(pkey = id, Target6hChange_Binary = response) %>% # Put back the original passenger IDs. No sorting has happened, so # everything still matches up. mutate(pkey = test$pkey) # Print preview of predictions print(head(prediction)) # Output data to Alteryx write.Alteryx(as.data.frame(prediction), 1) - Do whatever quick data prep is necessary. For example, if you are looking to trade on the top 3 predictions made, you could sort it by the prediction % descending and then only select rows 1-3 and use that information to initiate trades in the final step.
- Finally, your last step would be to connect to some sort of exchange or service in order to execute the trades. Typically you will be able to do this using an API provided by the exchange, and in some cases (like ours), there will be a developer that creates some sort of library that makes it easier to interact with the API, so if that's an option it's worth looking into it.
I hope this helps and makes sense, I shared what I felt like was useful and relevant to your question and I'm not sharing the workflow itself for now because I honestly don't think it would be very useful as nothing would run with all the database connections breaking, but if there is a demand for it I will give it some thought and try to make some form of template available. Let me know if you are still looking for more info, but I should definitely mention that it took a lot of time and effort to setup the infrastructure to get the data to flow into the database used for this post in close to live time, and although there have been some positive results from the predictive modeling itself, it has not yet translated into a consistently positive trading strategy, so I would encourage you to do a project like this to enhance your learning rather than assuming it will actually end up working.
Let me know if you have any questions!
Ricky
