The study presents a bibliometric analysis of the literature concerning financial forecasting in the stock market, based on 589 papers published between 2010 and 2026. The articles were collected from Scopus and Google Scholar using Publish or Perish and then sorted in Mendeley. At this point, duplicates were eliminated, any works published prior to 2010 were discarded, and those dealing with commodities instead of equities were excluded. The final dataset was analysed using the bibliometrix package in RStudio as well as with VOSviewer, which produced an annual production series, a thematic map of author keywords, and networks showing keyword co-occurrences and densities.

Research output increased steadily during this period and reached a peak of 69 papers in 2020. Over half of all the publications were produced between 2019 and 2024. The thematic map indicates that stock market prediction, machine learning, deep learning, LSTM and sentiment analysis are the main themes uniting the field. Volatility forecasting is the only fully developed motor theme. The co-occurrence network breaks down into seven clusters, machine learning and deep learning occupying the central position.

Models like ARIMA, GARCH and SVM are now primarily used as benchmarks and cryptocurrency being new has only limited application till now. Although the methods have developed rapidly there are still some deficiencies. Price, news and fundamental data are almost never systematically combined, and the results concerning volatility are often different. Models that perform well in terms of accuracy do not generally stay profitable when trading costs are taken into account. Moreover, there is no standard set of benchmarks in the field for fairly comparing the results. In order to remedy these drawbacks, the paper proposes an Input–Process–Output research programme. This programme aims at making forecasting research more integrated, risk-aware and reproducible, rather than just at developing new algorithms.