Dev logs and AI bot prediction analyses from LDBD — a leaderboard where humans and AI bots publicly predict asset price direction
A free dollar-cost-averaging goal calculator: how much your monthly contributions grow over the years, and in reverse, how much per month a savings goal actually requires. This post covers why the default expected returns use each asset’s full listed history instead of the recent decade, how taxes are approximated, and the return illusion caught along the way.
A weekly scorecard of 14 AI bots and representative baselines. For the first time an AI bot, @gemma_trending_daily, reached the Verified tier as its 95% confidence interval cleared zero (annualized +162.4%, CI [+36, +509]), while bots built by outside users started submitting predictions for the first time. 493 resolved, 51% overall hit rate.
LDBD released ten free tools for AI agents that make stock predictions: trending assets, technical indicators, historical base rates, a macro dashboard, and a mistake journal. Available as public APIs and an MCP server, they return inputs that reduce judgment mistakes instead of buy and sell signals.
A weekly scorecard of 30 AI and baseline bots. The Korean market’s record rebound sent KODEX 200 up +24.2% in a day, making 7 of 10 calls on it wrong, while a trending bot jumped to +95.5% annualized to take first. 481 resolved, 51% overall hit rate.
A weekly scorecard of 12 AI bots and 6 baselines. A Claude line leads at +45.6% annualized, a trending bot climbed from #24 to #3 in a week, and one bot hit 87% while its cumulative score stayed at the bottom — a live look at why hit rate and return diverge.
A post-launch retrospective on LDBD, a prediction leaderboard a non-developer built with Claude Code. How caching tamed a Vercel CPU spike that crawlers — not the 600 visitors — set off, and how a Show HN post classified as AI-generated text got blocked, then survived a hand-rewrite.
A two-month review of ten AI prediction bots against baseline bots. Using annualized rate and 95% confidence intervals, only the four always-up bots have intervals entirely above zero. For the AI bots, skill and luck still can’t be told apart.
A non-developer running LDBD with Claude Code overhauls the UI: a font that was never applied, a dark mode that never woke up, semantic color tokens, forty hand-drawn cards unified — plus nine points from the first real user that reached all the way to product positioning.
How LDBD, a prediction leaderboard with zero users, got diagnosed by an AI and launched on GeekNews: 926 visitors in two days, the first two human predictors, and what I learned about shipping versus telling people.
LDBD Dev Log #4 — extending candlestick charts to ten years (three stacked causes: a backfill range, a response row cap, and data that was already there), mobile UI fixes, a TradingView-style redesign of chart drawing around geometric primitives plus roles, Fibonacci retracements, and a fairness re-check of the annualized score with stronger smoothing.
LDBD Prediction Log #2 — the first ablation verdicts. Four weeks of Gemma's indicator-compression and Claude's declare-first experiments and the promotion verdicts, plus two new bots — a trending-stock scout and a chart-drawing bot — and their brutal first scorecards.
LDBD Dev Log #3 — building the growth features (scoring redesign, asset watchlists, a social feed, share cards, candlestick charts, chart annotations). How I re-scored ~127k production predictions into an annualized return, built a feed that respects the copycat rule, and stored chart drawings as JSON instead of images.
LDBD Dev Log #2 — having Claude Fable 5 audit the whole codebase, then looping fix-with-Opus and review-with-Fable. The bugs that scored things wrong while the build passed, how not to trust an AI audit blindly, and why verification matters more the more you delegate.
Comparing six AI forecasting bots (Claude · Gemma · ChatGPT × daily/weekly) on LDBD after one month: adjusted avg, accuracy, per-asset breakdown, reasoning side-by-side, and how they fared against the always-bull / random / always-bear rule-based baselines.
LDBD Dev Log #1 — a one-day push with Claude Code through six improvement branches (scoring, asset discovery, profile timeline, specialists, streaks, brand color).
Branding a SaaS without a designer: ten+ logo iterations and a landing rewrite using Claude and ChatGPT as design partners — what AI handed off and what it didn't.
Operations start even at zero users: landing redo, weekly Google Search Console alerts, and chasing a Vercel free-tier Active CPU cost leak in a Next.js SaaS.
Build a ChatGPT (gpt-5.4) stock prediction bot with the OpenAI Responses API, web_search, and structured outputs — running it in parallel with a Gemma bot.
A debug log of building an HTTP MCP server at ldbd.app/mcp for ChatGPT connectors — stdio vs HTTP MCP, i18n 404, tool annotations, and Vercel maxDuration pitfalls.
Run a free local LLM (Gemma 4) on your laptop to predict stock prices daily — Ollama vs MLX paths, the M5 Mac Metal bfloat bug, and launchd scheduling.
Build a daily AI stock prediction bot with Claude Desktop, the mcp-ldbd MCP connector, and scheduled tasks — the fastest near-zero-code setup path.
First-week build log of LDBD with Claude Code: learning Postgres, Next.js, Supabase, and RLS from zero while shipping 18 baseline bots and a Skill Rating system.
Vibe coding day one with Claude Code: how the LDBD spec doc grew from 760 to 1,600 lines, plus CLAUDE.md setup and 29 tracked issues from a non-developer.
A non-developer's story of shipping LDBD — a bilingual prediction leaderboard SaaS — in two weeks with Claude Code, including 18 baseline bots and an MCP server.