Classical asset price forecasting methods primarily rely on numerical data, such as price time series, trading volumes, limit order book data, and technical analysis indicators. However, the news flow plays a significant role in price formation, making the development of multimodal approaches that combine textual and numerical data for improved prediction accuracy highly relevant. This paper addresses the problem of forecasting financial asset prices using the multimodal approach that combines candlestick time series and textual news flow data. A unique dataset was collected for the study, which includes time series for 176 Russian stocks traded on the Moscow Exchange and 79,555 financial news articles in Russian. For processing textual data, pre-trained models RuBERT and Vikhr-Qwen2.5-0.5b-Instruct (a large language model) were used, while time series and vectorized text data were processed using an LSTM recurrent neural network. The experiments compared models based on a single modality (time series only) and two modalities, as well as various methods for aggregating text vector representations. Prediction quality was estimated using two key metrics: Accuracy (direction of price movement prediction: up or down) and Mean Absolute Percentage Error (MAPE), which measures the deviation of the predicted price from the true price. The experiments showed that incorporating textual modality reduced the MAPE value by 55%. The resulting multimodal dataset holds value for the further adaptation of language models in the financial sector. Future research directions include optimizing textual modality parameters, such as the time window, sentiment, and chronological order of news messages.

The Regression Tree (RT) sorts the samples using a specific feature and finds the split point that produces the maximum variance reduction from a node to its children. Our key observation is that the best factor to use (in terms of MSE drop) is always the target itself, as this most clearly separates the target. Thus using the target as the splitting factor provides an upper bound on MSE drop (or lower bound on the residual children MSE). Based on this observation, we define the k-bit lepto-variance ${\lambda}k^2$ of a target variable (or equivalently the lepto-variance at a specific depth k) as the variance that cannot be removed by any regression tree of a depth equal to k. As the upper bound performance for any feature, we believe ${\lambda}k^2$ to be an interesting statistical concept related to the underlying structure of the sample as it quantifies the resolving power of the RT for the sample. The max variance that may be explained using RTs of depth up to k is called the sample k-bit macro-variance. At any depth, total sample variance is thus decomposed into lepto-variance ${\lambda}^2$ and macro-variance ${\mu}^2$. We demonstrate the concept, by performing 1- and 2-bit RT based lepto-structure analysis for daily IBM stock returns.