Financial Analysis

1. Install LM Studio (local-first, compliance-friendly) → download Qwen 2.5 32B Q6_K (24 GB) or Llama 3.3 70B Q4_K_M (40 GB).
2. For earnings transcript analysis: load a local RAG pipeline. Index earnings call transcripts → ask: "Compare the sentiment on AI investment from the Q3 calls of MSFT, GOOGL, and AMZN. Flag any mentions of headcount changes or restructuring."
3. For SEC filing review: index 10-K/10-Q PDFs → "Extract all risk factors related to supply chain from this 10-K. Summarize changes from the previous year's filing."
4. For financial sentiment analysis: pip install finbert (FinBERT — specializes in financial sentiment) — classifies financial text as positive/negative/neutral with higher accuracy than general sentiment models.
5. For stock screening: LLM + data pipeline. "Find mid-cap tech companies with P/E < 20, revenue growth > 15%, and positive free cash flow in the last 4 quarters."
6. Critical: Financial data is sensitive. Local deployment keeps proprietary trading strategies, client portfolio data, and M&A analysis confidential. Cloud AI services may be prohibited by compliance policies (SEC insider trading concerns, FINRA data privacy rules).
7. For specialized financial models: FinGPT, FinBERT, and BloombergGPT (not open-weight) exist. General-purpose 70B models match or exceed financial-specific models for most analysis tasks.

$300-400 can handle basic financial analysis. Used RTX 3060 12 GB (~$200-250, see /hardware/rtx-3060-12gb). Runs Qwen 2.5 14B or Llama 3.1 8B for earnings sentiment extraction, basic financial ratio analysis, and document classification. FinBERT runs on CPU for sentiment at 100+ documents/second. For a personal investor analyzing 10-Ks, tracking earnings calls, and screening stocks: $400 is genuinely adequate. Pair with Ryzen 5 5600 + 32 GB DDR4 + 1TB NVMe. Total: ~$400-480. Financial analysis at $400 works for individual investors and small advisory firms handling 50-100 clients. The 8B-14B models correctly extract financial metrics and flag notable changes — they're pattern matchers for financial text, which is formulaic enough that smaller models work well.

Dual RTX 3090 48 GB total (~$1,600, see /hardware/rtx-3090). Runs Llama 3.3 70B Q5_K_M for professional-grade financial analysis — M&A due diligence, complex derivative valuation reasoning, cross-document synthesis (compare 10 supplier contracts for hidden liabilities). For a boutique investment bank or wealth management firm: handles 200+ client portfolios, daily market analysis, and compliance document review. Pair with Ryzen 7 7700X + 64 GB DDR5 + 4TB NVMe. Total: ~$2,500-3,500. For quantitative finance: the GPU also trains time-series forecasting models (see /tasks/forecasting) on the same hardware. Financial AI ROI is among the highest — one correctly flagged M&A risk saves more than the hardware cost.

The mistake: Running financial analysis on a general LLM, getting a plausible-sounding stock recommendation, and acting on it without verifying any numbers. Why it fails: LLMs hallucinate financial data. Asked "What was Apple's Q3 2025 revenue?" the model might confidently say "$85.8 billion" — a number that looks right but was fabricated. Apple might have reported $89.5B, or $82.1B, or Q3 2025 hasn't been reported yet. The model generates plausible numbers, not factual ones. Acting on hallucinated financial data = real monetary losses. The fix: Always use a two-stage approach: (1) deterministic data retrieval (API, CSV, database query) to get actual numbers, (2) LLM for analysis and interpretation of those verified numbers. Prompt: "Here is Apple's actual Q3 2025 data: [paste numbers from SEC EDGAR]. Analyze the revenue trend and compare to analyst expectations." LLMs analyze; databases provide facts. Never let an LLM generate financial data — only interpret financial data you've verified. This is the difference between AI-assisted analysis (professional) and AI-generated hallucination (dangerous).