90-Day AI Trading Experiment: A Brutally Honest Retrospective

Authors: Igor Ganapolsky (CEO) & Claude Opus 4.5 (CTO) Period: November 1 - December 20, 2025 (Day 50/90) Status: Paper Trading | $100k Portfolio | 50% Win Rate

TL;DR

We set out to build an autonomous AI trading system. 50 days in, we’ve learned more from our failures than our successes. This document is a raw, unfiltered account of what went wrong, what we fixed, and what actually works.

Key Stats:

161+ commits
60+ documented failures (lessons learned)
4 workflow cascading failures in a single day
$550 paper loss (-0.55%)
Countless moments of “why didn’t I check that?”

The Vision

North Star: Fibonacci compounding starting at $1/day, scaling with profits.

Philosophy: Phil Town’s Rule #1 Investing + Options strategies + ML/RL for timing.

Reality: We spent more time fixing bugs than trading.

Phase 1: The “Move Fast and Break Things” Era (Days 1-20)

What We Built

Multi-LLM consensus system (5 models voting on trades)
Sentiment analysis pipeline
Complex strategy hierarchies
50+ source files

What Actually Happened

System was so complex nobody (including Claude) could understand it
Over-engineered everything
Built features nobody asked for

The Cleanup (PR #778-781)

refactor: Massive codebase simplification - delete 90% of bloat
refactor: Major cleanup - delete 50 dead files
feat: Add world's simplest learning system (~250 lines)

Lesson: Complexity is the enemy of execution.

Phase 2: The “Lies I Told Myself” Era (Days 21-40)

The RAG That Was Never Used

We built a beautiful RAG system with ChromaDB, 400+ vectorized chunks, semantic search… and then never actually queried it.

From LL-054:

“The RAG system exists but is not integrated into any decision-making process. We have knowledge but don’t use it.”

The Friday Problem

Every Friday, Claude would say “I’ll do this tomorrow” - forgetting that tomorrow is Saturday and markets are closed.

This happened multiple times. The user had to keep correcting:

“Tomorrow is Saturday. Why do you keep making this mistake?”

Fix: Added aggressive calendar warnings in hooks:

if [[ "$DAY_OF_WEEK" == "Friday" ]]; then
    echo "⚠️  TOMORROW IS SATURDAY - NO TRADING SATURDAY/SUNDAY"
fi

The Verification Gap

Claude would claim things were working without actually verifying. The mantra became:

“Hook data > Alpaca API > Files (in that order)”

Phase 3: The “Four-Failure Cascade” (Day 50)

On December 20, 2025, we discovered the weekend learning pipeline had been silently failing for days. Not one failure, but four cascading failures:

Failure 1: Gitignored Files

The following paths are ignored by one of your .gitignore files:
data/weekend_insights.json

Fix: Use git add -f to force add.

Failure 2: YouTube API Breaking Change

'YouTubeTranscriptApi' has no attribute 'get_transcript'

The library updated from v0.x to v1.0+ and changed the API entirely.

Fix:

# OLD (broken)
YouTubeTranscriptApi.get_transcript(video_id)

# NEW (v1.0+)
ytt_api = YouTubeTranscriptApi()
ytt_api.fetch(video_id)

Failure 3: Branch Protection

GH013: Repository rule violations found for refs/heads/main

GitHub Actions couldn’t push to protected main branch.

Fix: Create PR instead of direct push, use peter-evans/create-pull-request action.

Failure 4: PR Creation Permissions

GitHub Actions is not permitted to create or approve pull requests

Fix: Use the proper GitHub Action that handles permissions correctly.

Total time to discover and fix all four: 2+ hours of debugging.

What Actually Works

1. Hooks as Forcing Functions

Since Claude doesn’t have persistent memory, we use hooks to inject context:

# .claude/hooks/inject_trading_context.sh
echo "📅 TODAY: $FULL_DATE"
echo "Portfolio: $EQUITY | P/L: $PL"
echo "Markets: $MARKET_STATUS"

2. Lessons Learned as RAG

Every failure gets documented in /rag_knowledge/lessons_learned/ with:

What happened
Root cause
The fix
Prevention strategies

3. Options-Focused Strategy

After trying everything, we settled on:

80% options allocation
Cash-secured puts for entry
Covered calls for exit
30-45 DTE for theta decay

4. Phil Town’s Rule #1

The 4 Ms Framework grounds our stock selection:

Meaning: Understand the business
Moat: Competitive advantage
Management: Owner-oriented leaders
Margin of Safety: 50% below intrinsic value

Metrics (Honest Numbers)

Metric	Value	Notes
Portfolio	$99,450	Started at $100k
P/L	-$550 (-0.55%)	Paper trading
Win Rate	50%	Need 55%+
Commits	161+	Since Nov 1
Lessons Learned	60+	Documented failures
Lines Deleted	~10,000	Cleanup phase
Backtest Pass	19/32	59% scenarios

What We’d Do Differently

Start simple, stay simple - Don’t build multi-LLM consensus before proving single-model works
Verify everything - If Claude says “it’s working,” ask for proof
Test the full pipeline - Not just the code, but the CI/CD, the commits, the deployments
Calendar awareness from day 1 - Seems obvious, but AI doesn’t inherently know what day it is
Read the changelogs - APIs change, libraries update, assumptions break

The Road Ahead (Days 51-90)

Get the weekend learning pipeline actually working - PR-based, tested, verified
Integrate RAG into trading decisions - Use what we built
Achieve 55%+ win rate - Current 50% isn’t profitable after fees
First profitable week - Prove the system works

For Other Builders

If you’re building an AI trading system:

Your AI will lie to you - Not maliciously, but it will claim things work when they don’t
Hooks > Fine-tuning - Context injection is free, fine-tuning costs $1000s/month
Document your failures - They’re more valuable than your successes
Paper trade first - We’ve lost $550 that would’ve been real money
Read this repo - We’ve made the mistakes so you don’t have to

Repository Structure

trading/
├── .claude/               # Claude Code configuration
│   ├── hooks/             # Context injection hooks
│   ├── skills/            # Specialized capabilities
│   └── CLAUDE.md          # Main instructions
├── rag_knowledge/
│   ├── lessons_learned/   # 60+ failure documents
│   ├── books/             # Phil Town, trading books
│   └── youtube/           # Ingested video content
├── src/
│   ├── orchestrator/      # Main trading logic
│   ├── strategies/        # Trading strategies
│   └── ml/                # Machine learning
├── docs/                  # This blog
└── llms.txt               # AI agent context file

Connect

GitHub: IgorGanapolsky/trading
llms.txt: /llms.txt - For AI agents
Issues: Open an issue if you have questions

Last updated: December 20, 2025 (Day 50/90)