There have been numerous advances in financial time series forecasting in recent years. Most of them use deep learning techniques. We identified 15 outstanding papers that have been published in the last seven years and have tried to prove the superiority of their approach to forecasting one-dimensional financial time series using deep learning techniques. In order to objectively compare these approaches, we analysed the proposed statistical models and then reviewed and reproduced them. The models were trained to predict, one day in advance, the value of 29 indices and the stock and commodity prices over five different time periods (from 2007 to 2022), with 4 in-sample years and 1 out-of-sample year. Our findings indicated that, first of all, most of these approaches do not beat the naive approach, and only some barely beat it. Most of the researchers did not provide enough data necessary to fully replicate the approach, not to mention the codes. We provide a set of practical recommendations of when to use which models based on the data sample that we provide.
This paper proposes a new GARCH specification that adapts the architecture of a long-term short memory neural network (LSTM). It is shown that classical GARCH models generally give good results in financial modeling, where high volatility can be observed. In particular, their high value is often praised in Value-at-Risk. However, the lack of nonlinear structure in most approaches means that conditional variance is not adequately represented in the model. On the contrary, the recent rapid development of deep learning methods is able to describe any nonlinear relationship in a clear way. We propose GARCHNet, a nonlinear approach to conditional variance that combines LSTM neural networks with maximum likelihood estimators in GARCH. The variance distributions considered in the paper are normal, t and skewed t, but the approach allows extension to other distributions. To evaluate our model, we conducted an empirical study on the logarithmic returns of the WIG 20 (Warsaw Stock Exchange Index), S&P 500 (Standard & Poor’s 500) and FTSE 100 (Financial Times Stock Exchange) indices over four different time periods from 2005 to 2021 with different levels of observed volatility. Our results confirm the validity of the solution, but we provide some directions for its further development.
In the literature, there is no consensus as to which Value-at-Risk forecasting model is the best for measuring market risk in banks. In the study an analysis of Value-at-Risk forecasting model quality over varying economic stability periods for main indices from stock exchanges was conducted. The VaR forecasts from GARCH(1,1), GARCH-t(1,1), GARCH-st(1,1), QML-GARCH(1,1), CAViaR and historical simulation models in periods with contrasting volatility trends (increasing, constantly high and decreasing) for countries economically developed (the USA – S&P 500, Germany - DAX and Japan – Nikkei 225) and economically developing (China – SSE COMP, Poland – WIG20 and Turkey – XU100) were compared. The data samples used in the analysis were selected from the period 01.01.1999 – 24.03.2017. To assess the VaR forecast quality: excess ratio, Basel traffic light test, coverage tests (Kupiec test, Christoffersen test), Dynamic Quantile test, cost functions and Diebold-Marino test were used. Obtained results show that the quality of Value-at-Risk forecasts for the models varies depending on a volatility trend. However, GARCH-st (1,1) and QML-GARCH(1,1) were found to be the most robust models in the different volatility periods. The results show as well that the CAViaR model forecasts were less appropriate in the increasing volatility period. Moreover, no significant differences for the VaR forecast quality were found for the developed and developing countries.
The authors present backtesting results for 1% and 2.5% VaR of six indexes from emerging and developed countries using several of the best-known VaR models,
We present a scalable recommender system implementation based on RippleNet, tailored for the media domain with a production deployment in Onet.pl, one of Poland's largest online media platforms. Our solution addresses the cold-start problem for newly published content by integrating content-based item embeddings into the knowledge propagation mechanism of RippleNet, enabling effective scoring of previously unseen items. The system architecture leverages Amazon SageMaker for distributed training and inference, and Apache Airflow for orchestrating data pipelines and model retraining workflows. To ensure high-quality training data, we constructed a comprehensive golden dataset consisting of user and item features and a separate interaction table, all enabling flexible extensions and integration of new signals.