Frequently Asked Questions
General Questions
Q: What do I have to learn to use Monument?
A: There is a learning curve with Monument, but it is much less than any other AI or machine learning curve that requires coding. A great place to start learning is on our Youtube channel, which shows easy-to-use examples to get started.
Visit Our Youtube Channel Visit Our Blog Book A Free Zoom Walkthrough
Q: How secure is Monument?
A: Monument is completely self-contained. You don't need an internet connection to use Monument. You can check this by turning off your internet and running algorithms. It all runs locally. You can't find this anywhere else. There is no risk of people snooping in on your work.
Q: What is the max data size Monument can support?
A: Monument supports data sets up to 1 terabyte.
Cryptocurrency Trading Questions
Q: How can Monument used for cryptocurrency?
A: Monument was created for algorithmic forecasting and makes algorithmic trading accessible for everyday cryptocurrency traders.
Q: How does Monument help me trade?
A: When you select a pair to analyze, Monument will let you know where various algorithms anticipate the price to be in the future. To do this, drag and drop an algorithm over the data.
Q: How is Monument different from crypto bots?
A: Monument allows you to build your own bots that no one else has access to. You get to choose the algorithms, parameters, and time intervals.
Q: Is there backtesting?
A: There is currently no back testing, but we are planning on adding it in the future.
Q: What is backtesting?
A: You would build a model that, as an example, says my target price is $9000 and every time it is below $8500 buy bitcoin and every time it is above $9500 sell bitcoin. Then you would simulate what would have been your profits if you did that strategy for the past month. To give a sense of what you might expect if you started doing that into the future.
Q: What is the trading time frame Monument works with?
A: Monument works with any time interval. It all depends on the data that is being uploaded.
Q: Where does the pair data come from?
A: We have direct integration with Binance.
Q: How do I know which algorithm to use?
A: It takes some practice to understand which algorithms are appropriate for which datasets. For example, certain algorithms do better when there is higher volatility in the training data. Whereas other algorithms do better when they are large, slow macro price movements.
Q: What are algorithm parameters used for?
A: Sometimes it's necessary to adjust the algorithm parameters in the interface to improve the results. Generally, a good workflow is to try several algorithms on the same data and see if they are agreeing generally, or if there is really wide disagreement.
Q: What makes an algorithm more likely to be accurate than another?
A: You want to pay attention to the "validation error rate", which is reported in the bottom left info box in the model interface. You want that number to be as negative as possible. The more negative that number is, the more likely the model is to be accurate. You want to experiment with algorithms and parameters until you find the lowest validation error rate.
Watch this video for more information
Q: What is a good gauge for the "validation error rate"? How far negative should the number be?
A: The more negative the number the better the model is. Models can always get better. The judgement on whether the validation error rate is good or bad comes with experience. That being said, it doesn't have to be negative per se, it could just be the lowest positive number you can get to. 10 is better than 100, -1 is better than 10, etc..
Q: What is the chart next to the "validation error rate"?
A: That chart shows you how the error rate improves with each training round. The x axis is each round. If the line is flat that means that it's not improving much each time. If there is a slope downwards, it means you might get better performance if you increase the number of training rounds.
Q: Can I predict any crypto pair or only certain ones?
A: A model is only ever as good as the input data. Sometimes the input data does not have predictive power.
For example, when the BTC halving happened, it affected the price of BTC. But this effect is not really detectable in advance from just the input data. An algorithm would have likely guessed incorrectly because that price movement was driven more by news than some "natural" pattern in the historical data.
Also, there are obviously whales in the market that move the price to run pump and dump. It would be difficult to detect that with algorithms.
Q: Is this a foolproof way of predicting prices?
A: No. Imagine that tomorrow some whale is going to do something crazy to move the market, but you trained your algorithm on all the data up to today. The validation error rate could be -100 (which is quite good) and you are feeling good about it. But the whale (an event external to the data) moves the market and it turns out the model was way off. All algorithms do is try to detect patterns in numbers. They cannot anticipate or account for anything that is outside of the numbers.
Q: Can I use Monument for FOREX trading as well?
A: Yes! Monument works with any kind of time series data. This includes thousands of things, for example: crypto, stock prices, forex, sea surface temperature, website visitors, number of times it rains per week, and many more.
Q: Why did we build Monument?
A: Most of our team is from the traditional finance world. We built Monument because we realized that a lot of the work that data scientists do is procedural and not actually worth the cost. Monument can automatically handle a lot of what data scientists need to use code to solve.
The goal with Monument was to make it so easy to get started with AI that everyone can hunt for the ideal parameters on their own. Instead of having to deal with a lot of complicated code just to even get started in the first place.
Q: What is the "windowing period" parameter?
A: The window period reflects the minimum amount of historical data that a model is trained against. You can train against more data by extending the window period or changing the windowing method to "expanding," which will cause the window length to increase over time. Data within the window is then sliced by the "lookback period" to train the model. Increasing the lookback period within the window will increase the length of the model, but since you only have a fixed amount of data available within the window, increasing the lookback period will reduce the number of training examples available to train the model.
Q: How can G-Boost and Entropy be used with missing data points?
A: G-Boost and Entropy Boost both interpolate missing data points with Gaussian Process interpolation.
Q: What does overfitting mean?
A: Overfitting is when you have trained an algorithm to fit better on the training set than possible on the out-of-sample set.
Q: What is Gaussian Boosting?
A: Creates an estimate of the range of your data. It creates a high point and a low point in your existing data.
If you want to know if the next point is really high or low compared to the previous data. It can estimate really well.
Q: What algorithm should I use for currency modeling?
A: The most overhyped algorithm right now is LSTM. We recommend spending the most time on Gradient Boosting (LGBM, G-Boost, Entropy Boost, G-DyBM). This isn't to say other algorithms can't make predictions, but we've had the most success using these algorithms.
LSTMs are hard to train, but they work really well for some problems. For financial issues, they are not very useful.
We recommend starting out with a basic AR model to see what is the most naive assumption and what it gets you. Then, a Gaussian Dynamic Boltzmann Machine (G-DyBM) because they train really fast and have better underlying assumptions. After, a dynamic linear model or time varying auto regressions. Then, go to boosting. Finally, go to LSTMs.
Q: What if I have more questions?
A: We provide direct support to explain things if they are confusing. We are always interested to hear feature requests from users. We can add things to the product if they will help you get more value from Monument.
Linux Installation Questions
Q: What dependencies do I need if I am running on Linux?
A: You must install libomp5 (sudo apt install libomp5). After, extract it [cd] into the directory and type ./Monument. Let us know if you have trouble with this.
Q: I can't access some of my files using Monument, how do I fix this?
A: This happens when some files have been written as sudo and cannot be updated when running the app as an ordinary user.
Assuming this is the case then this might have happened if the application was installed with sudo, but then run as the user. On Linux by default, if a file is created with sudo then it cannot be updated by the user and this gives a permission error.
If so, it can be fixed by changing the owner of the Monument files from root to the user.
In a terminal:
sudo chown -R user:user /home/YOURUSERNAME/Desktop/monument
sudo chown -R user:user /home/YOURUSERNAME/.config/Monument
sudo chown -R user:user /home/YOURUSERNAME/.config/MonumentAI
and similarly for other files/directories created when run as sudo. The user should first confirm that these directories do exist.
There may be other files not specific to the monument app that are owned by root. If necessary, to ensure everything under the user's home directory is owned by the user:
sudo chown -R user:user /home/YOURUSERNAME
Visit Our Youtube Channel Visit Our Blog Book A Free Zoom Walkthrough