Does open interest predict where coins go next? Traders swear by it: rising OI means new money and the move should run, falling OI means shorts covering and it should fade. I tested 477 Binance perps, Feb 2022 to Jun 2026.

A chart pattern with no reason behind it is just noise that hasn't failed yet. Most traders find the pattern first, then go looking for the reason. Quants do it backwards. I built one idea that way and tested it. It showed up in 142 out of 142 sessions. Here's the process 👇 Coffee is $3 downtown and $8 at the airport. Same coffee. The only difference is the buyers can't leave. That's what an edge is: someone on the other side paying more than they should, for a reason they can't avoid. There are only three: 1. They're forced. Liquidation engines market-selling at any price. Desks rolling futures before expiry. 2. They're buying insurance. Traders overpaying for puts and vol into FOMC or CPI. 3. They're in a hurry. Every market order pays the spread and the taker fee to get filled now. None of those are patterns. They're people with a reason. Run the same test on your indicators. A moving average is arithmetic on old prices, and nobody's bonus depends on it. VWAP is the price big orders get graded against, so a desk buying size all day is actively trying to beat it. TWAP is the same idea on a clock: slice the order evenly through the day. One is math. The other has a person with an incentive behind it. Once you have a reason, turn it into a question: when X happens, is Y different from the baseline? Condition, outcome, baseline. If you can't fill in all three, you don't have an idea yet. Then, before you touch the data, write down what would prove you wrong: - It disappears after costs - It only works in one year - It shows up just as strongly when your reason isn't there Then run the test. If any of those show up, the idea is dead, no matter how good the reason sounded. If none do, you might actually have something. Here's the one I tested. Big orders get sliced on a clock, so is there more volume at the top of every minute? On 142 sessions of index futures data, the first second of each minute had about 1.8x the volume of a normal second. Every single day. Not a pattern. A schedule. Crypto runs 24/7 and a lot of size moves through TWAP bots. Run the same test on BTC perps and tell me what you find. Most ideas die at the prove-me-wrong step. That's the process working. The idea isn't the edge. Killing bad ideas fast is.

How I use heatmaps to avoid overfitting: Split your data in two. One part you optimize on (in-sample). The other you lock away and don't touch until the end (out-of-sample). Run your parameter grid on IS. Then, instead of picking the single best cell, average each cell with its neighbors and pick the best average. Why? One great cell surrounded by bad ones is usually luck. A cell whose whole neighborhood performs well is much more likely to hold up. Then run the same grid on OOS and check whether that area still looks good. If it does, you might have a real edge. If it falls apart, you were fitting noise, and it's better to learn that now than with real money. This is what a robust strategy's heatmap should look like:

There are only 4 ways to actually make money in markets: - Provide a service (market making) - Harvest a risk premium (ERP, index beta) - Know something first (informational edge) - Exploit how the market is mechanically wired (structural edge) That's it. That's the list. If you can't put your strategy in one of those buckets, you don't have an edge. Prove me wrong if you can.

Survivorship bias is probably the most common way people fool themselves in a backtest. Especially in crypto. If you're testing on coins that are listed on Binance today, you're testing on survivors. Dozens of tickers have been delisted over the past few years (dead volume, regulation, or the project just quietly stopped existing). Your strategy would have bought those. It might have held them all the way down. But they never show up in your equity curve, because they're not in the data you pulled. Same trap with "top 20 by market cap." Today's top 20 is the list of coins that made it. The 2021 top 20 looked nothing like it. Plenty of those names aren't even top 100 now. Run your backtest on the current list and you're selecting winners you had no way of identifying in advance. The result is always the same: the backtest looks better than the strategy is. The fix is easy. Use data that includes delisted assets. This image is the classic version of the same mistake. WWII analysts mapped bullet holes on returning bombers and wanted to armor the red areas. Abraham Wald pointed out the obvious thing nobody had said: those are the places a plane can get hit and still come home. The planes hit in the engines and cockpit weren't in the dataset. They never landed... Your delisted coins are the missing planes.

Added a new strategy to my portfolio yesterday with a 2.24 sharpe, 0.52 correlation to the book. It's a market neutral, made up of cross-sectional idiosyncratic momentum + a secondary signal. Parameters were chosen in-sample. Both signals held up clean out-of-sample. Higher correlation than i'd like, but it clears the bar. The test isn't whether correlation is low though, it's whether sharpe beats correlation × book sharpe. Book's around 3, so the bar was 1.56. it cleared. took the portfolio from ~3.0 to ~3.1. Both signals have a real economic rationale for why they should make money, and it holds across almost any parameter choice -> not a curve-fit That's the whole decision rule for adding a strategy and almost nobody uses it

been running this systematic portfolio on hyperliquid for 2 months now. live's ahead of the backtest, partly because i modeled fees pessimistically. added a new sleeve today 2.22 sharpe standalone, 0.52 correlation to the existing book. might write that one up

Your Sharpe ceiling is set by correlation. Not by how many strategies you run. Correlation here just means how often your strategies win and lose together. At correlation 0.25: 10 strategies gives you 1.75x 40 strategies gives you 1.93x The ceiling is 2.00x and nothing gets you past it. Now hold it at 10 strategies and drop correlation to 0.10. You get 2.29x. A quarter of the strategies, more Sharpe. Building strategies is easy. Building ones that are uncorrelated is the hard part.