Temerity Holdings @temeritytrd

My first market-making model: HMM regime detection, KDE price distribution by regime, bid-ask surface as output. I half knew how to vibe code — built the whole thing by copy/paste from ChatGPT. Deployed without a finished backtest. Unsuccessful. Fail faster with Claude Code!

@Robot_Wealth "The bloke... knows the fish sit behind that rock when the water’s up, and move to the eddy below the bend when it drops. He catches fish every time... Because he understands the river." The understanding here is observational, not explanatory. This bloke could be a data miner

@Robot_Wealth Agree that overfitting is a common pitfall and data mining / vibe quanting are great ways to get there. But does an edge really need to be fully explainable? Markets are complex dynamic systems whose behavior does not always fit nicely into constrained explanations

@zostaff This is great stuff thank you for sharing

@zostaff Why not operate in logit space for the full range of prices? Switching at the boundaries seems cumbersome

@zostaff The original AS model assumes a "large and constant" arrival rate of market orders. Is the Hawkes process your way of compensating for sporadic event driven prediction market order flow?

@zostaff Is the "Limit order fill intensities decay exponentially in distance from mid" assumption valid for prediction markets? Low liquidity, thin book markets especially. Could behave far more erratically

Working hypothesis: when an MLB game gets shortened or delayed by rain, baseball totals markets reprice inefficiently. Fewer innings, wet conditions = fewer runs. I suspect retail knows this and overreacts.

Step 3: score each survivor. Effective spread × daily volume × √(depth-penetration ratio) × log1p(taker pressure). Rank descending, review the top 20 by hand. The algorithm finds, then I review manually.

Three metrics go into the score. Effective spread: VWAP-walked round-trip cost at 50 contracts. Depth-penetration ratio: effective spread ÷ top-of-book — how much wider the book gets at real size. Taker pressure: daily volume ÷ total resting depth.

I'm building a market-making bot for Kalshi (first iteration of many). There are 10s of thousands of open markets on the platform (57k+ when I ran this analysis). The first problem is figuring out which ones are actually worth quoting. Here's the funnel I built. A thread.

HFT owns every edge that lives in milliseconds. The interpretation edge — which markets reprice after a refinery fire, and in what order — lives in seconds. The positioning edge around a data release lives in days. Neither is HFT territory. That’s where slow money competes.

Green = top-of-book. Blue = spread at 50 contracts. Red = 200 contracts. The markets that blow out with size are expensive for takers to trade and candidates for makers, because the spread is actually there to capture.

A market where effective spread is still 2¢ at 200+ contracts has deep, competitive liquidity — not much room for a small maker to earn. A market where it jumps to 20¢ has a thin book. Thin books are where a maker's quotes can earn real spread.

Prediction market exchanges advertise top-of-book spread — the most flattering liquidity number they could show you. But spread isn't a single number. It's a spectrum, and where that spectrum sits is what matters for a market maker. A thread.

@craigsuperstar @robinhanson "median" doesn't say much about the 3 estimates individually lol