Predicting the Unpredictable: Neural Networks Learn to Filter Market Noise
Scientists at the Interdisciplinary Scientific and Educational School of Moscow State University “Mozg, kognitivnye sistemy, iskusstvennyy intellekt” (Brain, Cognitive Systems, Artificial Intelligence) have improved the accuracy of financial time-series forecasts by 1.3–1.7 times compared with a baseline model.
To achieve this, they combined a neural network with fuzzy logic methods and tested the approach on BTC–USD quotes across three time intervals. The development could help create more robust methods for analyzing market data. The research paper describing the results was published in the Communications in Computer and Information Science series.
Financial markets are chaos disguised as patterns. Even the most advanced neural network runs into the same fundamental problem: market noise. Market noise consists of sporadic, short-term price fluctuations and random information signals that do not reflect an asset’s underlying value or correspond to an objective price trend. Moscow scientists have found a way to work around that obstacle. They proposed a hybrid approach that combines the conventional long short-term memory (LSTM) recurrent neural network with fuzzy logic.
Mathematics vs. Market Noise
The core of the development is a modified LSTM architecture. Fuzzy logic is incorporated at two levels: during data preprocessing and inside the network itself, where activation functions are interpreted as membership functions for fuzzy sets.
Put simply, a standard neural network tries to identify rigid mathematical patterns, breaking down when it encounters anomalies and panic selling. The MSU hybrid model accounts for uncertainty, operating more like an experienced analyst who works in terms of “likely” and “most likely.”
The results from tests on historical BTC–USD quotes are notable. Forecast accuracy increased 1.3 times for minute-level data, 1.7 times for hourly data and 1.5 times for daily data compared with a conventional LSTM. The trade-off is longer model training time because of the additional parameters.
A “Magic Crystal” for Traders or a Research Prototype
Does this mean AI that can reliably predict the price of bitcoin is just around the corner? Not so fast. That is not what the study shows. It would be premature to call this theoretical research a ready-to-use fintech product. The experiment covered only one asset, and the comparison was made against a baseline architecture rather than the most advanced model available.
For everyday consumers, there is no direct impact yet. Over time, however, algorithms like this could underpin bank risk-assessment systems and corporate analytics. In finance, experience shows that the model itself is only part of the equation. Choosing the right features and metrics and thoroughly validating the system on test data are just as critical.
Russia’s AI Context: From News to AutoML
In recent years, Russian researchers and businesses have made major advances in predictive analytics. Consider some of the key milestones: in 2023, HSE University and VTB developed a model that predicts stock-price movements based on news sentiment; in 2024, neural networks demonstrated an advantage over conventional econometric methods in forecasting inflation; and in 2025, HSE researchers trained AI to anticipate stock-market crises, while the Moscow Exchange launched its own MOEX_AutoML toolkit for business forecasting. The MSU hybrid approach fits into this ecosystem by targeting a specific challenge: noisy data, which is common in both Russian and global financial markets.
Scale vs. Accuracy
In 2024–2025, the global IT industry moved toward scale, building huge general-purpose models, such as GPT for text, that train on enormous datasets and attempt to handle many tasks at once rather than focusing on specialized solutions.
Amazon released Chronos, while Google introduced TimesFM 2.5 – foundation models designed to cover time series at massive scale, turning numerical data into something resembling language tokens. Much like language models such as GPT and Gemini are trained on text, these models are trained on millions of different time sequences, ranging from product sales to stock-market quotes and sensor readings.
Russian researchers are taking a different path. Instead of building a giant general-purpose model, they are improving the existing LSTM architecture, which is smaller and simpler, by adding fuzzy logic to handle uncertainty. It is a “point solution” for a specific problem – noise in financial data – rather than an attempt to cover every time series in the world.
Export Potential and What Comes Next
The next step for the researchers is to move beyond cryptocurrencies. The algorithm needs to be validated on stocks, bonds, oil and macroeconomic indicators. Ideally, the hybrid model should not be used as a standalone “predictor,” but as one component of a system that combines multiple algorithms.
If the technology passes further validation and is packaged as a software API, it could offer export potential. As of 2025, Russian fintech companies were steadily expanding their presence in CIS, Middle Eastern and Global South markets. A hybrid neural network that is resilient to the volatility of emerging markets could become a useful product for banks and exchanges abroad.
The MSU development is not a tool for financial speculators but a foundational approach to advancing predictive analytics. And it may be that the future of financial AI lies with hybrid systems that can acknowledge their own uncertainty.
