Classical and modern financial NLP

6.1.1 Tokenisation, embeddings; financial dictionaries (Loughran-McDonald), FinBERT.
6.1.2 Sources: news, 10-K, 10-Q and 8-K (EDGAR), CNMV material disclosures (hechos relevantes), earnings call transcripts, social media.
6.1.3 Event studies with text signals; precise time alignment (publication timestamps); topic modelling.

LLMs for research and trading

6.2.1 How LLMs work (context, temperature, hallucinations); commercial APIs versus local models (Ollama, vLLM).
6.2.2 Structured prompting and JSON outputs validated with Pydantic.
6.2.3 Information extraction: guidance, KPIs and tone of conference calls; LLM sentiment versus FinBERT.
6.2.4 Lightweight fine-tuning (LoRA, QLoRA) for financial tasks.
6.2.5 Look-ahead bias in LLMs: the model “knows the future” of its training corpus; backtests with cut-off dates and point-in-time models.

6.3.1 Embeddings and vector databases (FAISS, pgvector, Chroma, Qdrant).
6.3.2 Chunking financial documents: tables, footnotes, XBRL.
6.3.3 RAG over annual reports, ETF prospectuses and KIDs and in-house research; evaluation (faithfulness, citation, coverage).

6.4.1 Agents with tools (tool calling), ReAct, planning.
6.4.2 Frameworks: LangGraph, LlamaIndex, provider SDK agents; MCP (Model Context Protocol) to connect data and brokers.
6.4.3 TradingAgents-style multi-agent architectures: fundamental, sentiment, news and technical analysts, bull and bear debate, trader and risk manager.
6.4.4 Evidence and limits: in LiveTradeBench (50 live sessions, 21 LLMs), the Spearman correlation between LMArena score and cumulative return on equities was only 0.054 (−0.38 on Polymarket). GPT-4.1 gained more than 6% on equities and lost more than 30% on Polymarket. AlphaForgeBench documents “drastically different” actions across runs of the same model, even with deterministic decoding. Deep learning models that are strong in backtest degrade in live trading.
6.4.5 Guardrails: the agent proposes and the deterministic risk engine decides; hard limits, human approval, audit log.
6.4.6 Research agents: hypothesis generation, assisted alpha mining, code and backtest review.

Fig. — LiveTradeBench: correlation between LMArena score and the returns of 21 LLMs

6.5.1 Coding assistants (Copilot, Claude Code, Codex, Cursor): a tests-first workflow.
6.5.2 Automatic review of backtests for biases; migrations (pandas 2 → 3, backtrader → NautilusTrader).
6.5.3 Risks: plausible but incorrect code, invented dependencies, leaked secrets.

6.6.1 When it is “big”: tick data, L2 and L3 books, options (OPRA), global universes.
6.6.2 Columnar formats (Parquet, Arrow); lazy and streaming Polars; out-of-core DuckDB.
6.6.3 Dask and Spark (PySpark) for clusters; Ray to parallelise experiments.
6.6.4 Time series databases: TimescaleDB, QuestDB, ClickHouse, ArcticDB; kdb+ as professional background.
6.6.5 Cloud storage (S3, GCS), lakehouse (Delta, Iceberg) and GPU (RAPIDS, cuDF) as background.

6.7.1 Landscape and market size: Neudata figures and caveats about consultancy reports with widely varying estimates.
6.7.2 Types: web scraping (prices, jobs, reviews), web and app traffic, card transactions, receipts, geolocation, satellite, shipping (AIS), weather, patents, public data, on-chain.
6.7.3 Responsible web scraping: requests and httpx, BeautifulSoup, Playwright; robots.txt, terms of use, GDPR.
6.7.4 Providers and due diligence (Neudata, Nasdaq Data Link, Eagle Alpha): point-in-time history, panel biases, legality (inside information, personal data).
6.7.5 From data to signal: mapping to tickers (entity resolution), sales nowcasting, signal backtest, alpha decay.
6.7.6 Case studies: satellite imagery with computer vision (CNN, ViT) and supply chain graphs.