Decentralized cryptocurrencies have gained a lot of attention over the last decade. Bitcoin was introduced as the first cryptocurrency to allow direct online payments without relying on centralized financial entities. The use of Bitcoin has vastly grown as a financial asset rather than just a tool for online payments.

A lot of cryptocurrencies have been created since 2011 with Bitcoin dominating the cryptocurrencies’ market. With plenty of cryptocurrencies being used as financial assets and with millions of trades being executed through different exchange services, cryptocurrencies are open to trading problems and challenges similar to those encountered in the financial domain.

Price and trend prediction, volatility prediction, portfolio construction and fraud detection are some examples related to trading. In addition, there are other challenges that are specific to the domain of cryptocurrencies such as mining, cybersecurity, anonymity and privacy.

In attempts to solve, many apply Machine Learning, the subcategory of AI into finance. The AI is excellent in pattern recognition. So the idea is, if the AI can see patterns in the price data — the chart — it can also tell which direction is the price likely to move next. Let's explore the application of artificial intelligence techniques to address these challenges for cryptocurrencies with their vast amount of daily transactions, trades and news that are beyond human capabilities to analyze and learn from.

Price Prediction/ Forecasting

The basic flow of most of the work done in this area starts with the collection of time-series data for different variables involved. These variables include some market (economic and financial) and Social Media (Sentiment) indicators.

Analysis of the data and relationships between different variables and the cryptocurrency price are then deduced. A supervised machine learning technique is used to learn a model from data which can then be used for prediction. Using the history of different variables makes price prediction a time-series prediction task.

It can be modeled as a regression problem to predict the closing price based on a set of indicators. It can also be modeled as a classification problem to predict if there will be a rise/fall or no change in the price of a coin by encoding the cryptocurrency price time series output variable in terms of rise and fall.

Volatility Prediction

Volatility of cryptocurrencies is mainly caused by their decentralized nature making their prices uncontrollable by any organization or government. In extension, cryptocurrencies can be considered as being traded in a free market where the price is solely determined by the supply and demand, however, there are other factors.

A cryptocurrency price range could be estimated if volatility can be predicted or estimated for a day or a week based on historical data. Generalized autoregressive conditional heteroskedasticity (GARCH), which is a time-series statistical model, is used for modeling volatility.

Automated Trading

Trading bots are software products or cloud services that offer what is called ‘‘algorithmic trading’’ as they automatically analyze market actions and indicators, offer strategies to maximize trader’s gains and improve ROI. They can aggregate historical market data, calculate indicators, simulate order execution and even can be set up to execute strategies without customer supervision or initiative.

Some bots even use natural language processing techniques to communicate with the customer in a more natural and friendly way. In the design of these trading bots, many algorithms and techniques similar to those used for price and volatility prediction mentioned above are used to maximize the profit and develop a strategy with maximum return.

Often these differ in the number of exchanges they support and the features they offer. Additionally, they can offer portfolio construction and optimization to find an optimal weighing of financial assets which might include Bitcoin, other cryptocurrencies and other traditional financial assets like stocks and bonds. This optimization aims at maximizing the return while minimizing the variance of the return.

Botsfolio.com is a textbook example demonstrating the above traits. Learn more about it here.

Fraud Detection

The usage of Bitcoin and other cryptocurrencies in facilitating illegal activities is a big concern, as it affects the stability and the trust in cryptocurrencies. Cryptocurrencies are known to attract cybercriminals for their pseudo-anonymity and for being operated outside the regulation regimes of governments and banks.

However, regulators are continuously trying to enforce know-your-customer (KYC) and anti-money laundering (AML) laws for exchanges and escrow services. There are different types of scams and criminal activities that can occur in cryptocurrencies, such as digital theft, hacking, phishing, Ponzi-schemes, pump-and-dump schemes, purchasing illegal drugs and money laundering in the black market.

Fraud detection is based on detecting anomalies and suspicious behaviour in the transactions and trades history, especially that Bitcoin transactions are transparently recorded on the blockchain public ledger.

But without recorded incidents or examples for different fraud activities, trimmed k-means and k-means clustering based on features from the transactions graph in a semi-supervised way can be used to detect fraudulent activity in the Bitcoin transactions network.

Based on the clustering labels for outliers, some tweaked classification models can be used to understand the relation between the labels and predictor variables. Random forest achieves the best precision.

Anonymity and Privacy

Privacy and anonymity are two necessary aspects for online financial trading. Anonymity is mostly favored by preferred by privacy-savvy people who want to keep their identities and transactions anonymous and private. Privacy means protecting the data of transacting users including the traded amount, the transacting parties, their balances and the timing of the transaction.

Trying to reveal the identities of Bitcoin users and linking their Bitcoin addresses and trades usually rely on using public data information from social media or other publicly available data in a process called ‘‘de-anonymization’’.

It is either based on heuristics to link this data to the blockchain transactions, or it can be based on AI techniques. Deanonymization has been approached using AI in two ways; clustering or classification.

Cryptocurrency Mining

The mining process has the disadvantage of high electricity consumption used by mining pools for participating in the POW computations. Only one miner succeeds to add a block of transactions, while other mining pools are left with the expenses of huge energy costs. This disadvantage threatens the decentralization of the cryptocurrency and makes it susceptible to 

monopolization, especially when the block reward will vanish over time due to Bitcoin block reward halving.

Game-theory analysis is being research for use to improve the block reward allocation in mining to keep the blockchain secure. AI techniques can be employed by mining pools to choose which cryptocurrency to mine and which mining pool to join in order to reduce the electricity consumption and increase their profit based on historical data.

Security

Despite the security and privacy attributes that exist in blockchain-based cryptocurrencies, there are several security threats that are facing the cryptocurrency ecosystem. They can be classified as attacks on the distributed network, mining process attacks, double spending and transaction malleability attacks. There are also client-side security attacks and privacy threats to wallet, exchange or escrow services.

DDoS attack detection in Bitcoin-related services (e.g. mining pools, currency exchanges, e-Wallet, gambling services) have been studied using the Game-theory. A cryptocurrency network is more vulnerable to attack if selfish miners controlled more than a certain threshold of computational hash power. Selfish mining behavior has been modeled by using Markov-models and analyzed.

The above mentioned research for security analysis either use supervised machine learning models or game-theory techniques.