Not any trading software

We have built a high-performance trading software using Machine Learning and Deep Learning Artificial Intelligence (AI) technologies.

The trading software learns the best timing to place a trade, how big or small the position size should be, and when to exit a trade. To us, position sizing is as important as trade entries and exits.

Prediction models are very specific, so currently, the trading software is fine-tuned specifically for cryptocurrency pairs.

Our main hypothesis is that the financial markets have all the information 'fairly' priced in but there are still gaps where the data could be processed. The old traditional statistical methods like linear regression or moving averages may have worked in the past but we feel that there are more advanced techniques like Machine Learning and Deep Learning that will be able to push the performance much further.

In data science, when we are analyzing time-series data, how we process the data is critical. In the crypto space, data is widely accessible. But before the data could be utilized, we will need to clean the data before our Machine Learning and Deep Learning quant model begins analyzing them. Deep learning techniques are necessary as they are able to refine unstructured data to build the most accurate model.

Our quant model is built to identify market structures left by the Market Makers. It will be able to predict how Market Makers move in the market, detect any inventory level imbalances, and how the Market Makers will potentially skew the order books to restore this balance. We aggregate this data and we will then take a position in the market to exploit this move.

Building this low-frequency, high-capacity quant model started way back in January 2017 and it was first created in Python and C++. In mid-2017, the painstaking and time-consuming process of migrating from Python and C++ to Pine Script began, and it took us about a year to complete the entire migration process. Forward testing of the pre-production quant model started in mid-2018 for another year before it finally went live.

Understanding financial data and the cryptocurrency market
‍First, we need to know and understand the financial data that is available in the crypto space. We are predominantly looking at off-chain and on-chain data. On-chain data refers to data that is recorded on the blockchain which includes information about transactions and it can provide insight into the supply and demand for a particular asset as well as the level of activity on the network. Off-chain data, on the other hand, refers to data that is not recorded on the blockchain i.e. market conditions, exchange order book data, and other relevant factors that may affect the price of the asset. Both on-chain and off-chain data can be useful for making informed decisions.

‍Machine Learning and Deep Learning academic papers
‍The next step is to establish the baseline in terms of the academic papers, where the cutting-edge model is sitting, and how they designed those models. It took us a while to do literature reviews on these papers and a lot of trial and error. 70% of these models we tried to replicate didn’t work or gave mixed results. On the other hand, 30% of them have some best practices and what we did was combine all of those models together.

‍Running trading algorithm analyses, analyzing results, data preprocessing, and designing the Machine Learning model 
‍We started to run some baseline analyses of how we can use the data and analyze the results. We then began building the data processing pipeline and started writing the model in C++ and eventually Python because of the readily available Machine Learning and Deep Learning libraries. We ended up building the model from the ground up and when developing a Deep Learning model, this process involves designing the Neural Network Architecture as well. It is essentially layers and layers of how these things are being processed, so we’ve spent a lot of time developing the optimal architecture that gives the highest accuracy.

‍Ensembling
‍Lastly, we started to train the model and what happens is that it may give 20 to 30 different models. We then used a popular technique called Ensembling. What Ensembling does is combining all of the different models into one supermodel.

Keras and Tensorflow are the two Deep Learning Libraries that we used to train the quant model and process time-series data such as cryptocurrency price feeds. Scikit, Seaborn, and Scipy are the traditional statistical libraries that we used for data analysis. Finally, XGBoost is a feature engineering tool that trains and refines the parameters of a feature that affects the model. For example, we feed data in terms of price, transaction, volume, etc. and if the data are highly correlated, we will then focus on those datasets.

We separate the data into test sets and train sets. For example, we broke up the Historical Data into a specific timeframe and assigned it to the train sets, and we will then use the test sets to see if we can predict the results.

Weights are assigned via a correlation matrix. The quant model will prioritize which one correlates more. Then, from a Deep Learning perspective, it assigns the weight to the predictive input on its own in order to push for a higher Profit Factor. We don't tell it what to do; the model will figure it out itself.

The formula to calculate Profit Factor is Gross Profit divided by Gross Loss and, it is simply the gross profits made for each trade as a ratio of the gross loss. If it equals 1, then you are barely making any profits. World-class quant models range between 2 to 2.5. With Deep Learning technology, we are able to push it above that range.

We have not re-tuned our quant model ever since 2018 which was during the pre-production stage. So we will only start re-tuning the model when we see that the model produces a Profit Factor of less than two, and this is where we will use a slightly different architecture for the model.

The quant model predominantly focuses on high-performing trades vs. high-frequency trades. The model trades on a longer timeframe and a typical trade would last anywhere between 2 to 6 weeks or it will take trades anywhere between 15 to 18 in a year.

Yes, definitely, and it is one of the main objectives. Since we are only making 15 to 18 trades per year, we can perform such trades as we do not want to be in the market or create the market structure. Inversely, if you are executing high-frequency trades, you create the market structure, which is not ideal. 

Our quant model main objective is to optimize the Profit Factor, a standard used commonly by traders to measure the effectiveness of a strategy. 

We don't focus on Sharpe Ratio as our quant model trades on a low-frequency basis, between 15 to 18 trades per year. We can smoothen up the equity curve or increase the Sharpe Ratio by making more trades, but it will decrease the Profit Factor.

The percentage profitability varies between the different crypto-asset trading pairs. On average, it is about 40-50%. We tried to optimize this figure, but unfortunately, it pulled down the Profit Factor.

Our Net Profit remains high because when our quant model predicts the right moves, it is able to position size the trades optimally while keeping the losses low when it predicts the wrong moves.

Risk managing a trade is often difficult. However, the quant model uses machine learning and deep learning to mitigate some of these risks including position sizing the trade, assessing the leverage the model takes, and ensuring the whole process can be computed in a short time. Thus, it is able to achieve a consistent profit factor. 

There is no pre-defined Stop Loss within the model. From our forward-testing observation of how the model behaves, the model closes out its position when there is no longer an inventory risk held by others on the market.

Noisy data is common in any marketplace. This is where training the model becomes essential. Ultimately, each data is assigned a correlation weight and the model uses that to filter out the noise.

Yes, you can choose any crypto-asset trading pair, select a timeframe, change your initial capital, which is denominated in USD that you will be starting with, and adjust your equity risk per trade. Other conditions that you can customize but are not limited to are fees incurred per trade, slippage levels, Long entries only, Short entries only, or both Long and Short entries, to name a few. 

It is the financial risk involved in holding equity in a particular investment. Select a lower equity risk for trades that you intend to take with lower leverage. For zero leverage, you will have to use an Equity Risk of 0.1 and below.

Yes, you can backtest our trading software. You can choose your desired Start Date, Start Month and Start Year. It can be backtested across the duration of when the chosen cryptocurrency was listed in that particular exchange. For example, it can backtest BTCUSD over the last five years as long as BTCUSD is traded on that exchange.

In data science, overfitting can be solved by a wide range of techniques, particularly for time-series data. 

‍Weight decay is a regularization technique that adds a penalty to the loss function used to train the model. This penalty encourages the model to use smaller weights and helps to prevent overfitting.

‍Cross-validation is a technique for evaluating the performance of a model by dividing the data into a number of folds and training the model on a different subset of the data in each fold. This can help to prevent overfitting by ensuring that the model has not been trained too closely on the training data.

‍Early stopping is a technique that involves monitoring the performance of the model on a validation set as it is being trained, and stopping the training process when the model's performance begins to degrade. This can help to prevent overfitting by limiting the amount of training the model receives.

‍Simplifying the model, such as by reducing the number of features or the complexity of the model, can also help to prevent overfitting.

‍Ensembling involves training multiple models and combining their predictions to make a final prediction. This can help to reduce overfitting by averaging out the errors of individual models.

‍Granger Causality Test is a technique to determine the accuracy of the quant model. It is a statistical hypothesis test for determining whether one-time series data is useful in forecasting another. XGBoost and Scipy are used to get the correlation metrics across all these input data and how it affects certain things. 

Overall, we are using a combination of these techniques to prevent overfitting and improve the generalization ability of our quant model. This may get too technical but it’s something we pay close attention to. Another way is to do forward testing – please reach out to us at info@almanacinc.ai for more info. 

Yes, it is running live and we have been forward-testing our pre-production model since mid-2018. We have officially published the trading software in June 2019 and the trading signals are based in real-time.

We will only trade high liquidity assets for example, BTC, ETH, and BNB, to name a few. We will stick with the settings that give the highest Profit Factor and it should also comprise a smooth and healthy-looking equity curve. 

The quant model excels in both Bear and Bull markets, and it is justified by the Profit Factor of the chosen trading pair. The higher the number, the higher the probability of having a profitable trading pair.

The execution engine is non-proprietary. We connect to existing platforms over API or the trader can trade manually when the software sends the trading signal. The exchanges that we trade on have liquidity of up to notional USD 1T per year or 200,000 BTC daily.

Of course not! Net Profit and Max Drawdown are something that we cannot predict or decide. It is entirely market-driven. Past performances do not dictate future results. The main objective of the quant model is to optimize the Profit Factor of the chosen crypto-asset trading pair. 

Yes, we do. However, do note that we work with business entities only. No private individuals allowed. Please reach out to us at info@almanacinc.ai for more info. 

Yes, we can. However, do note that we work with business entities only. No private individuals allowed. Please reach out to us at info@almanacinc.ai for more info. 

We are not a fund and we do not manage any investments. We are a crypto-native software provider and our main objective is to extract alpha from the crypto space. The client will have 100% custody of the trading capital. The trades will be executed via the client’s Institutional sub-account. Thus, we will only work with business entities that have Institutional accounts on their preferred cryptocurrency exchanges.

We are based out of Singapore and Vietnam. Vietnam is where our developers reside, and Singapore is where we carry out business and commercial matters.