RAG systems Skills for data scientist in investment banking: What to Learn in 2026

AI is changing the data scientist role in investment banking in a very specific way: the job is moving from building isolated models to designing systems that can retrieve, ground, explain, and audit outputs against firm data and market context. If you cannot build or evaluate RAG pipelines over research, filings, policies, client notes, and internal knowledge bases, you will get boxed into reporting work while the higher-value AI work moves elsewhere.

The 5 Skills That Matter Most

1. •

   Document ingestion and financial text normalization

   Investment banking data is messy: PDFs, OCR’d scans, pitch books, earnings transcripts, credit memos, and compliance docs all show up in different formats. You need to know how to extract structure from this material reliably, because RAG is only as good as the chunks you feed it.

   Focus on parsing tables, preserving section hierarchy, handling footnotes, and normalizing entities like issuer names, tickers, dates, and deal terms. For a data scientist in investment banking, this matters because bad ingestion creates hallucinations that look plausible in front of bankers and risk teams.

2. •

   Retrieval design for high-precision search

   Generic semantic search is not enough in banking. You need hybrid retrieval: keyword + vector + metadata filters + reranking, so the system can find the right clause in a 200-page credit agreement or the right precedent transaction in a noisy deal archive.

   Learn how to tune chunk size, overlap, embeddings, BM25 weighting, and rerankers for finance-specific queries. This skill matters because most banking use cases fail when retrieval returns “close enough” instead of the exact policy paragraph or disclosure language needed for decision support.

3. •

   LLM evaluation and grounding

   Bank users do not care if a demo looks smart; they care if it is correct under audit. You need to measure retrieval accuracy, answer faithfulness, citation quality, and refusal behavior on out-of-scope questions.

   Build eval sets from real workflows: due diligence Q&A, KYC policy lookup, research summarization, or earnings call extraction. A strong data scientist in investment banking should be able to say not just “the model works,” but “it achieves 92% citation accuracy on our internal test set and fails safely when evidence is missing.”

4. •

   Secure deployment and governance

   Banking AI lives inside strict controls: data residency, access control, retention rules, model logging, and approval workflows. If you cannot design around entitlements and audit trails, your system will never make it past legal or technology risk review.

   Learn how to integrate document-level permissions into retrieval so users only see what they are allowed to see. This matters more than model choice because one access-control bug can kill an entire AI program.

5. •

   Workflow integration with banker-facing tools

   The best RAG system does not live in a notebook. It sits inside tools bankers already use: internal portals, SharePoint-like repositories, deal rooms, Slack/Teams bots with guardrails, or analyst workbenches.

   You need enough product thinking to embed retrieval into real workflows like preparing management presentations or answering diligence questions. For a data scientist in investment banking, this is the difference between being seen as an experimenter and being trusted with production systems.

Where to Learn

• •

  DeepLearning.AI — Retrieval-Augmented Generation (RAG) course

  • •Best for getting the core architecture straight: chunking, embeddings, retrievers, reranking.
  • •Good first step if you want practical grounding in 1–2 weeks.
• •

  Hugging Face Course

  • •Useful for transformers basics, embedding models, tokenization limits, and fine-tuning concepts.
  • •Spend 1–2 weeks on the sections relevant to text classification and retrieval.
• •

  OpenAI Cookbook

  • •Strong reference for building production-style RAG pipelines, evals, function calling patterns, and structured outputs.
  • •Use it as a working handbook while building your first internal prototype.
• •

  Full Stack Deep Learning

  • •Good for deployment thinking: monitoring, evaluation loops, reliability tradeoffs.
  • •Spend 1 week on the MLOps sections that map to regulated environments.
• •

  Book: Designing Machine Learning Systems by Chip Huyen

  • •Best single book for understanding why model performance is only one piece of the system.
  • •Read it alongside your first production RAG project over 2–3 weeks.

A realistic timeline:

• •Weeks 1–2: document parsing + embeddings + basic retrieval
• •Weeks 3–4: hybrid search + reranking + prompt grounding
• •Weeks 5–6: evaluation harness + citations + failure analysis
• •Weeks 7–8: access control + deployment patterns + workflow integration

How to Prove It

• •

  Internal research assistant with citations

  • •Build a RAG app over equity research notes or sector primers that answers questions with source citations.
  • •Show precision on named entities like companies, guidance numbers, catalysts, and valuation metrics.
• •

  Credit agreement / policy clause finder

  • •Create a tool that retrieves exact clauses from legal or compliance documents with metadata filters by entity type or jurisdiction.
  • •This proves you understand high-stakes retrieval where “close enough” is useless.
• •

  Earnings call summarizer with evidence tracing

  • •Ingest transcripts and generate summaries tied back to speaker turns and timestamps.
  • •Add evaluation for factual consistency so users can verify claims quickly.
• •

  Deal room Q&A assistant with permissions

  • •Build a prototype where users can ask questions across diligence docs but only see documents they are entitled to access.
  • •This demonstrates both RAG engineering and governance awareness.

What NOT to Learn

• •

  Toy chatbot frameworks without retrieval control

  • •If it cannot handle citations, permissions, or evaluation rigorously enough for banking use cases like diligence or research lookup, it is not helping your career.
• •

  Over-focusing on prompt engineering

  • •Prompt tricks do not fix bad ingestion or weak retrieval.
  • •In investment banking AI work in particular, system design beats clever wording every time.
• •

  Training large models from scratch

  • •That is not where most value sits for a data scientist in investment banking.
  • •Your edge comes from building trusted systems over proprietary content, not burning months on foundation-model experiments you cannot deploy internally.

Keep learning

• •The complete AI Agents Roadmap — my full 8-step breakdown
• •Free: The AI Agent Starter Kit — PDF checklist + starter code
• •Work with me — I build AI for banks and insurance companies

By Cyprian Aarons, AI Consultant at Topiax.