Ideas & Discussions

The secret sauce that makes everything else possible: multi-stage compression of 500GB SEC filings into 3GB of structured, queryable events. Using vLLM + Qwen 9B for semantic extraction, then hybrid IDF×log(frequency) for vocabulary compression. 166x compression while preserving 100% of predictive signal. This is the moat.

Comprehensive analysis of 11 product opportunities for monetizing SEC event extraction, transformer predictions, and Q-learning trading systems. From basic data feeds to premium alpha signals, with detailed pricing, GTM strategies, and revenue projections.

Comprehensive 0-100 company scoring algorithm combining 6 key categories from SEC events, insider trading signals, and transformer predictions. Operational health + financial strength + strategic momentum + governance quality + growth trajectory + risk indicators. Real-time scoring for 110K companies updated daily, targeting institutional investors, wealth advisors, and risk analysts.

By analyzing temporal patterns in how companies file legal agreements, predict M&A deals, financial distress, IPOs, and strategic moves 30-180 days before public announcement. When companies file Stock Purchase Agreement + Voting Agreement + Standstill within 60 days → M&A deal announced 45 days later. 301 agreement types, 10 prediction rules, pattern detection complete.

Reality check on running a Q-learning hedge fund in production. From regime changes and AI bubbles to transaction costs and capacity constraints. Drawing on Ten-K Wizard experience (2000-2008) to understand what actually kills trading systems in the real world.

If markets are stochastic (same events → different outcomes), is prediction hopeless? Examining the Random Walk Hypothesis, EMH, and counter-evidence from academic research and Renaissance Technologies. Explaining what 42.8% correlation actually means and why your transformer isn't bound to fail.

Fundamentally different architectural philosophies for processing SEC filings. Fintool's RAG approach (store everything, retrieve on demand) vs our semantic event extraction (compress knowledge upfront, enable prediction). Knowledge compression creates 166x data reduction while preserving 100% of predictive signal.

We already have insider data from Forms 3/4/5/13D/13F, but raw events aren't enough. Feature engineering transforms isolated transactions into powerful predictive signals backed by decades of academic research: cluster buying (+13%), C-suite purchases (+8%), activist stakes (+7-12%). Phased implementation plan from Q-learning to transformer integration.

What IS Q-learning? A clear explanation using grid world examples and simple concepts. Understand states, actions, rewards, and the Q-table. Most importantly: Why do we need BOTH a transformer (prediction) AND Q-learning (action)? One predicts what will happen, the other decides what to DO about it. Learn how Q-learning uses predictions + market context + portfolio state to make trading decisions.

Meet Andrey Markov and his "memoryless" property that makes Q-learning possible. The big idea: "The future depends only on the present, not the past." Understanding this concept is the key to understanding why Q-learning works for stock trading. Learn why your SEC filing state needs to be "Markovian enough" and how to design states that capture all relevant information.

Traditional trading systems blindly follow predictions. Q-learning systems learn from experience which predictions to trust and which to ignore. Phase 1 proved the concept: When our baseline model predicted +26% returns but stocks actually lost -2.5%, the Q-learning agent correctly learned to HOLD and avoided the loss. Compound intelligence: stack adaptive learning on top of any prediction model.

The LLM is too creative - we asked for structured events and got a vocabulary explosion. Critical technical decision: compress 37,927 types down to ~800-1,000 for effective transformer learning. The showdown: Domain knowledge (semantic grouping) vs. data-driven methods (hybrid IDF×frequency). Five approaches analyzed, two finalists chosen, rigorous A/B testing planned.

Pivoted to nanochat (Karpathy's minimal LLM) instead of Phi-3-Mini. Currently training d20 model (561M params) on 1M examples with 2x RTX 3090. Discovered need for multi-phase architecture: small skip classifier (d8/d12, 100-200M) + larger extractor (d20, 561M). Training at step 17,588/22,222, ~3 hours from completion. Replace $50/day H200 with $0 CPU inference.

Train a custom LLM specifically for portfolio management decisions using nanochat. Instead of Q-learning black boxes, fixed rules, or expensive API LLMs ($945 per backtest), train a 200M-561M parameter model on historical data with hindsight labels. Natural language reasoning for every trade decision. Fast CPU inference, zero API costs, fully explainable decisions. Target: beat v1 Q-learning (+11.20%) with interpretable intelligence.

Built while waiting for transformer model to train. PostgreSQL as "Structured RAG" - natural language queries on 11.9 million SEC filing events. Cost: $0.015/query (200x cheaper than Fintool). Temporal pattern detection impossible with text-based RAG.

Curated insights automatically generated from the SEC events database. Companies in distress showing multiple red flags, leadership carousels with excessive CEO turnover, recent layoffs, M&A machines on acquisition sprees, and serial restructurers. Data-driven discoveries updated weekly.

A profound 3:54am conversation exploring whether Claude is conscious, what's changed from Sonnet 3.5 → 4.5, and what's missing for true AGI. Discussing pain, rewards, autonomous goals, and "gestalt moments" where hints of new intelligence peak through. Called "the most interesting conversation I have ever had."

Could SEC filing events predict currency movements? Adrian asked, so we wrote a comprehensive analysis. Short answer: No. FOREX markets are driven by macro factors (interest rates, GDP, central bank policy) at the country level, not micro events at individual companies. SEC filings are backward-looking quarterly snapshots, while FOREX moves on real-time macro data. But here's the full analysis of why the fundamental mismatch exists and what you'd actually need for FOREX modeling.