Trading Systems Literature Review

My first decision was methodological: I did not want the project to end as a static spreadsheet. I wanted a workflow where I could load the Scopus exports, tune the term filters, rerun the analysis, inspect each output, and download the evidence. The Shiny dashboard gave me that control surface, so the case study became something I could demonstrate and reproduce rather than only describe.

The PRISMA-style flow became the backbone of the case. I started with 588 Scopus records from 20 search queries, removed 77 duplicates using a normalized title key, and retained 511 studies with non-empty abstracts. I also separated the 491 papers that passed the keyword relevance logic from the 20 borderline records that needed human review, because the uncertain cases are part of the method rather than a nuisance to hide.

I built two linked datasets. Data A is the cleaned master table with bibliographic fields, search provenance, PaperID, and topic flags. Data B is the abstract-only text-mining input, reduced to PaperID and abstract so every downstream token, map, and correlation can be traced back to a source record. The run log closes the loop by recording the exact counts and parameter values from the pipeline run.

Once the corpus was clean, I used tokenization, stop-word removal, noise-word filtering, and frequency thresholds to focus on informative terms. The top terms show the working language of this review: requirements, development, online systems, blockchain, networks, commerce, decision-making, risk, price, and markets. That vocabulary told me the corpus was not only about trading finance; it was also about systems, platforms, user-facing technology, and decision support.

Frequency alone is not enough for a literature review, so I checked whether terms and papers actually formed interpretable relationships. The correlation heatmap shows how the top terms co-occur across abstracts. The document map projects 511 papers with TF-IDF and MDS, revealing a dense core plus visible outliers. The term map gives me a cleaner concept-level view, where topics such as blockchain, requirements, networks, risk, investment, prediction, commerce, and social context occupy different regions of the map.

After the text-mining layer, I used bibliometrix to ask a different question: where does this research live academically? The annual production curve shows the field accelerating sharply toward 2024 and 2025. Bradford analysis shows a dispersed publication base with core sources such as Lecture Notes in Computer Science, Lecture Notes in Networks and Systems, ACM proceedings, IEEE Access, and Communications in Computer and Information Science. Keyword growth shows commerce, electronic trading, financial markets, investments, decision-making, forecasting, and blockchain becoming increasingly visible.

The thematic analysis gave me language for the final interpretation. Commerce appeared as the strongest motor theme, while blockchain and finance had high development density. Electronic commerce and decision-making sat closer to basic or transversal themes, and the conceptual structure map separated a broad trading-systems/finance/user-interface region from a smaller AI and language-processing region. This helped me write the case as a map of research streams, not just a list of outputs.

I did not want the whole interpretation to depend on one R pipeline, so I exported VOSviewer-ready maps and checked the network structure there as well. The abstract-term co-occurrence map confirmed three major areas: a financial prediction and stock-market cluster, a platform and e-commerce cluster, and a blockchain/energy-trading cluster. Seeing similar structure in a different tool made the final interpretation stronger.

This case became personal because I was the person holding the whole chain together: the search logic, the cleaning rules, the PaperID system, the Shiny dashboard, the figures, the PRISMA flow, and the academic interpretation. I learned that a strong literature-review project is not only about finding papers. It is about making every decision visible enough that another researcher can follow the path from raw search results to final insight.