Russian programmers have launched an online service in the United States , which predicts the change in stock prices, using neural networks and machine education. How accurate are the predictions?
Russian programmers Dmitry Lukovkin and Petr Permyakov offered a ready-made solution for traders, who do not want to get involved with technical indicators and financial performance of companies, - StocksNeural online service. He, according to the description, predicts the movement of prices for stock market instruments and issues recommendations on their purchase or sale, using deep machine learning (Deep Learning) and artificial neural networks.
Maximum term, for which StocksNeural provides a forecast, is five days. Wherein, as Dmitry Lukovkin says, theoretically, this algorithm is able to predict the change in the stock price for up to a month. According to him, forecast accuracy reaches 97% depending on the tool. This, however, does not mean, that the algorithm is not capable of making mistakes: for some securities, the number of "hits" is in the range of 30-40%.
Despite the origin of the founders, startup StocksNeural is registered in Wyoming, USA, and is focused primarily on global markets. Currently, the service provides information only on US stocks from the S indices&P500 и Dow Jones Industrial Average. In the future, the range of tools will expand - the creators of StocksNeural promise to include exchange-traded funds (ETF) and commodity futures.
Forecasts of prices of Russian shares are not yet included in the plans of StocksNeural - the Russian market is smaller in volume, than American, And there are far fewer active investors on it, explains Lukovkin. Currently, the StocksNeural service exists only in the web version, However, the company intends to release it in the form of a mobile application. "Clear, that the functionality of the mobile service should be more interactive, so that the user can not only see the forecast, but also to perform some action.", Lukovkin says.
How it works
StocksNeural is a site, which contains a list of analyzed stocks. Registration on it takes a couple of minutes: the user needs to specify an email, Password and bank card number. The subscription price is $49,95 per month or $395,95 in year, However, the former 15 days it can be used for free. Main, Do not forget to cancel the unsubscribe after the grace period, so that the service does not automatically write off money.
The site describes the method of price prediction. To simplify, then the company, using special filters, cleans quotes from noise - the influence of sharp fluctuations. Based on the resulting smoothed prices, the neural network identifies dependencies in the behavior of stocks over an interval of up to ten years and, starting from them, gives a forecast.
As the head of the big data laboratory of the Internet Initiatives Development Fund explained (IIDF) Anatoly Orlov, There is nothing supernatural in this algorithm. "If a standard predictive model makes a prediction based on those factors, which programmers have contributed to it, then a neural network is a model, which is capable of creating factors from raw data.", He says..
Schedule orange (cm. photo) reflects the actual smoothed value of prices, And the blue graph is predictive. When the forecast deviates by a certain amount from the actual price, The system gives a recommendation to buy or sell shares.
Investors can also see the accuracy of the execution of forecasts - it, according to Lukovkin, Calculated based on historical data. If the investor is only interested in stocks with the most accurate price prediction, it can go to the Members Area tab, where the securities are ranked according to this indicator. In the same tab, forecasts for price changes for five trading sessions in a row are given, and recommendations for this period.
Difficulties in forecasting
RBC decided to conduct an experiment with shares of IT companies, known as FANG: social network Facebook, Apple gadget maker, Netflix video streaming service and search engine Google. We were interested in the accuracy of the forecast of the prices of these securities for 17 august. Based on historical data, the accuracy of StocksNeural prediction for FANG shares was more than 90%.
Following the trading session 17 August it turned out, that the service guessed the target value of the stock only in one case . Google's share price reached its target price within a day. Lukovkin explains, What is the service, rather, shows a fair price for the paper and does not necessarily, so that the shares reach it during the day. "Here we need to look not at its meaning, and on the trend itself and on the recommendations for the promotion", - he stressed.
Waiting for a miracle
StocksNeural is far from the only service, trying to predict price movements using a specific algorithm. According to Lukovkin, when developing StocksNeural the experience of the IKnowFirst and StockCharts platforms was taken into account. BCS Marketing Director Ruslan Smirnov says, that there is now a growing interest in the financial industry in forecasting prices for stock market instruments using machine learning and neural networks.
Trader and asset manager of S.R. investment company. Solutions Roman Andreev, specializing in algorithmic trading, Notes, that market participants have repeatedly tried to create a stable algorithm based on neural networks, But none of these attempts led to the desired result. As soon as the market eliminates inefficiencies, allowing such algorithms to work, they quickly "burn out". According to Andreev, forecast accuracy at least in 90% seems like a fantasy result. The trader himself assesses the maximum possible accuracy of the forecast for the movement of the share price only in 70%.
Anatoly Orlov from IIDF believes, What is the algorithm, used by StocksNeural, can only work on a short time interval. On a longer horizon, the idea of predicting quotes, focusing on the analysis of data from the past, It is unlikely to justify itself, he says. "It's like trying to systematically predict 'black swans.', Orlov emphasized.
Yaroslav Podsevatkin, head of trading at Aton Investment Company, admits, What is theoretically the technology, the underlying StocksNeural, can work. However, the trader doubts, that it can be compared with technical or fundamental analysis. "Statement of high accuracy of the forecast, which provides this algorithm, more like a marketing ploy.", - he summed up
Read more at
