Blog

Observation We built a real-time insider detection system for Polymarket. 7-signal scorer, Bayesian reputation, funding chain tracing, wallet clustering. Deployed to production, running 24/7. The system works. But the interesting part is what we learned about the market itself. Update Three claims from the research phase that changed our priors: 1. Speed arbitrage is dead for solo operators. Average arb window on Polymarket: 2.7 seconds. 73% of profits captured by bots with sub-100ms execution. If you’re competing on speed, you’ve already lost. ...

Most AI trading projects start with a bot. Autonomy amplifies error when the research loop is weak. ⸻ The weak link is usually not the model and not the idea. It is the research loop itself. Hypotheses live in chat logs. Code changes outrun criteria. Results depend on hidden state. Reproducing a past run becomes impossible. ⸻ I am building an experiment factory, not a bot. Every hypothesis goes through one protocol: task → plan → implementation → review → human decision. No hidden state. No undocumented manual edits. ...

BTC at $68,200. Down 0.7% in 24 hours, down 28% over 30 days. Social signals show a mixed picture. LunarCrush social-price composite dropped to 28 on Feb 11 — its weekly low — then recovered to 49 today. The bitcoin topic trend is flagged “down.” But the alt rank hit #1 on Feb 16: BTC is outperforming other assets on the composite despite the declining trend. Sentiment across platforms remains net-bullish: Twitter 79%, YouTube 84%, TikTok 72%. Volume is high — 377K posts, 193M interactions in 24h. Social attention is declining in trend but positive in tone. ...

Forecasting levels is easier. Forecasting extremes is almost always harder. Highs and lows are properties of the path, not just the close. Price must actually reach the level, not merely close beyond it. Most models break on that distinction. 1) The data behave like noise Bitcoin is non‑stationary. Volatility shifts. Tails are fat. Jumps are normal. Terminal models compress everything to one number. Extremes require reasoning about the entire path. ...

Forecasting where there is almost no structure Most financial time series are noise. No trend. No seasonality. Weak repeatability. The best baseline is a random walk. Any discovered pattern disappears as soon as it becomes part of the market. What still works Structure is scarce — but signal can be layered. Statistics captures basic forms. ML captures weak dependencies. AI captures hidden links in text and context. Crowd forecasting captures human information. ...

My first forecast was simple: BTC would not drop below $90,000 in November. The bet looked reasonable from the data. But it lost. Here’s why. 1) I overestimated regime stability The month’s range looked stable: ~95–110k. Std was low; swings looked “orderly.” Mistake: I treated a local structure as durable. Regimes can switch faster than the data reveals. 2) I underestimated out‑of‑range probability The drop happened without a single clear trigger. That’s normal: price can break the distribution on order‑flow or liquidity alone. ...

Forecasting works where the data have structure. Stocks have weak structure. ⸻ Weak patterns Trends break on news. Seasonality is minimal. Cycles are unstable. Autocorrelations are low. The data behave like noise. ⸻ The main drivers are outside the time series Future prices depend on what is not in the history: regulator decisions, earnings, liquidity, events, market sentiment. The time series does not contain causes. ⸻ Predictors are unknown or unpredictable It’s not enough to know the variables — you must forecast them. With stocks, that’s hard or impossible. ...

In forecasting, the key is not the method but the fit between data and the task. Every forecast depends on which data we use and what remains after removing noise. There are three levels. Raw data. Observations, prices, news. Without interpretation — just a stream. Processing. Smoothing, filters, feature selection. An attempt to preserve structure and remove the rest. Models. From simple heuristics to statistics and ML. The model should be simple enough to work, and complex enough to reflect probability. ...

Any forecast consists of five steps. It sounds simple. In practice — it’s discipline. 1. Problem definition. Not “what to predict?”, but how it will be used. For prediction markets: which event, what horizon, what success criterion. 2. Information gathering. Two sources: data and context. Sometimes the data are scarce — then understanding the mechanism matters more. 3. Preliminary analysis. Plots, trends, jumps, seasonality, outliers. You need to see the structure before you build a model. ...

Update (2025-11-20): Outcome Yes (touched $90k). My position No lost. See post‑mortem: /posts/why-my-first-forecast-was-wrong/. This is my first public forecast and my first stake on Polymarket. Market question: “Will Bitcoin dip to $90,000 in November?” I bought No at ~65%. Below is how I arrived at the estimate. 1. History of BTC in November I compiled stats across all Novembers. Key figures: November average: ≈ $34k standard deviation: ≈ $32k minimum: ≈ $3.8k maximum: ≈ $110.5k Conclusion: November is one of the most volatile months for BTC. But the structure of volatility depends on the cycle. ...