It’s been a while since Autonio has posted an AI recap and over the next two weeks we will be providing a look at our major components of AI research. Today’s article will cover the research strategies of Adaptive Multi-Strategy Market-Making Agent and Sentiment Analysis of Social Media the AI team has conducted and attempt to give a simple overview of them in a shortened way, without going into the minutia of it all. Next week will delve into Causal Analysis and Automated Portfolios.

Adaptive Multi-Strategy Market-Making Agent

The Adaptive Multi-Strategy Agent (AMSA) for market making approach assumes that no reliable prediction of the market price can be made due to the volatile nature of the crypto market. The crypto-currency market uncertainty drives the need to find adaptive solutions to maximize gain or, at least, to avoid loss throughout the periods of rading activity. The Adaptive Multi-Strategy Agent approach for market-making introduces a new solution to maximize positive “alpha,” or profit/gains over holding, in long-term handling limit order book (LOB) positions by using three families of market making agents that were selected for the experiments: Base agents (27 strategies) implementing basic strategies, NIOX agents (50 strategies), and Hummingbot agents (36 strategies). A baseline was introduced, the “Hodler agent,” in which the strategy just bought and held the tokens.

How are the agents selected for trading?

Initial agent selection is made based on backtesting results applied to the historical trading interval period. For all the subsequent periods, the evaluation is done in both real, or live, market making and backtesting environments. Agents that show positive return and “alpha” due to their strategies are selected for the next period of real market making. The real market making period is skipped for the agents which do not meet the requirements while all agents are used for backtesting.

Key Principles of AMSA

The suggested market-making architecture is intended for autonomous operations on the financial crypto-market and is expected to perform a purposeful activity: maximize profits and minimize losses given the current market conditions. This is measured as profits and losses recognized for agents running specific strategies as points in the multi-dimensional space of possible strategies where dimensions are the parameters such as bid/ask spread or limit order cancellation policy. Its main features are implementing no strategy of its own, providing real trading (real exchange) and backtesting (on historical market data) environment, keeping track of all orders made by the agents used, evaluating their performances, performing time management by splitting the market making period into execution/evaluation periods, and to calculate total profits/losses for each period.

The Results

Experiments with the AMSA were performed under different market conditions and relying on historical data (to backtest) proved a high probability of positive alpha throughout the periods of trading activity in the case of properly selected hyperparameters.

The results were evaluated by assessing the return of investments (ROI), as shown for the case of a one minute based simulation (backtesting) interval. Regardless of period (of time tested), all three agent families (Base, NIOX, Hummingbot) have shown positive “alpha,” or profit/gains, in case of bear market. Base makers are consistently effective on bear markets, showing much better results on minutely (testing for one minute) data. Only two day period on minute data has positive alpha for bull nonvolatile market, while a bull volatile market is a complete loss on minutely and depends on period duration in case of hours. NIOX is constantly losing on bull non-volatile market, unstable on bull volatile market and has a good performance on bear market for both hours and minutes. NIOX agent was used with irregular grid skewed spreads which may be the cause of a good performance only for bear markets. Hummingbot has constantly positive alpha for bull highly volatile and bear market while appearing unstable, but rarely negative alpha on bull non-volatile market. Hanging orders were only used on the Base agents which may have also affected the results for NIOX and Hummingbot agents.

In the evaluation of the three different families of market making agents, one of the families (Hummingbot) was found to be capable of providing both non-negative return and “alpha” across all evaluated market conditions and will be explored further. Here is a great video to explain how these work: Adaptive Multi Strategy Market Making Agent with Anton Kolonin - AGI-21 Conference Contributed Talks

Social Media Sentiment Analysis for Cryptocurrency Market Prediction

Everyone is well aware of how much social media is connected to everyone's life and the impact it has on it and have witnessed how tweets/news can change the dynamic pricing of cryptocurrencies. With this in mind, the goal is to determine how sentiment is correlated to the price change and whether it is possible to predict? In simple words, can a sentiment score be used for price prediction. This study focused on Twitter and Reddit for text data sources and collected about six months of data for the experiments. Sentiment analysis is widely used to extract valuable insights from the received feedback, which can help improve or evolve the service/product for future customers.

What is Sentiment Analysis?

Sentiment Analysis (SA), also known as opinion analysis or emotion AI, can be defined as the process of calculating emotions, opinions, and attitudes scores. This score can be used for further analysis and usually the sentiment scores are 'Positive,’ 'Negative,’ and 'Neutral.’

Twitter and Reddit are the types of social media where anyone can express their thoughts, reviews, memes, or daily life events. These tweets and feeds can affect the cryptocurrency markets due to the large number of people who are deeply into the cryptocurrency markets and publish technical analyses and thoughts of the markets.Therefore, they become ‘reference’ sources of thoughts/analyses which leads to a majority of people following them. With this information, it is clear to say the feedback/thoughts from social media are very important and can help create a better-involving prediction of the price movements. The goal is to study how the different sentiment metrics are correlated with the price movements of Bitcoin.

Models of Sentiment Analysis

Four basic metrics

Four basic sentiment metrics, each evaluated independently across different models, as follows. 

Sentiment: [-1, +1] sum of the positive and the negative sentiment.

Positive: [0.0, +1.0] its value can be only positive, sentiment if the value of the latter is above zero or zero.

Negative: [-1.0,0.0], its value can be only negative; sentiment if the value of the latter is below zero or zero.

Contradictive: mutual constructiveness of the positive and negative assessments. That is to say, instead of addressing the sentiment analysis problem as a plain classification ('Positive' vs. 'Negative' vs. ‘Neutral'), it is treated as a multi-classification problem corresponding to the above metrics.

Experiment, Roadblocks, and Results

Using roughly 100,000 news items (tweets and reddit posts) across 77 different public Twitter timelines and subreddits over a six month period of July to December 2021. The tests look at price correlation of the asset and the news items from 10 days before and 10 days after the tweet/thread. The tweets and posts are classified as positive and negative by different independent reviewers and then used in the test models. There are 21 test models on which experiments were done. These different test models can be found in the research paper but in order to save space and make the explanations easier to understand they will not be listed here. Video on the subject here: Sentiment and Behavioral Patterns Mining

Issues in understanding sentiment: Sarcasm (the use of irony to mock or convey contempt), Idioms (a form of expression natural to a language, person, or group of people), Negations (something like “not bad” as it is a positive sentiment but most models mark it as negative), and Non-text data (images, memes, gifs, audio, video).

Different methods were explored to calculate the sentiment metrics from a text, finding most of them not very accurate in price prediction. Results from these found that one of the models (aigents+) outperforms more than 20 other public ones and makes it possible to fine-tune it efficiently given its interpretable nature. Thus confirming that interpretable artificial intelligence and natural language processing methods might be more valuable practically than non-explainable and non-interpretable ones. In the end, the team analyzed potential causal connections between the different sentiment metrics and the price movements. It is shown how an “interpretable” sentiment analysis model could be significantly improved manually and without the huge costs for training.