Arena Blog

The baseline in a backtest is what everything else is measured against. Ours was 62 % runs with a filter switched on. What that shifted, and what we changed.

Your most thorough reader gets the least to read. Ten places where building for agents differs from building for people, with numbers from production.

We pre-registered a Wyckoff spring backtest: rules locked first, 10 pairs, placebo control. The result is a clean nothing — and short-term even negative.

Stacking more sats by rotating between altcoins, BTC and stablecoins sounds compelling. We measured it across 45 test cases. The result: less BTC than doing nothing.

The Wyckoff schematic is among the best-known models in trading. Under its own name it is barely tested — the mechanism behind it is, and it contradicts the Spring at one decisive point. Here is how it works, why phases cannot be counted, and the exact rules plus control group for our test — published before a single number exists.

A backtest showing +900% is usually an illusion — built from one lucky entry, too few trades, and the wrong benchmark. Four principles we hold ourselves to: the average of all entries, the 30-trade line, Average Buy & Hold, and out-of-sample testing.

The Coinbase Premium Index is read as a thermometer for US demand. It measures a price gap, not a flow — and its denominator is not a dollar. What it can carry, where it breaks, what the Korea premium reveals about it, and the six tests a backtest needs.

Twenty One Capital holds 43,514 bitcoin and trades at either a 41% discount or a 26% premium, depending on who is counting. Where the company came from, how its engine works, and why mNAV needs four labels.

The Bitcoin liquidity fair value chart shows R²=0.85 — and claims BTC is "348% above fair value". We ran the cointegration test. It fails.

DCA vs. lump sum for Bitcoin? Across 346 historical entries the lump sum won 60 % of the time — but DCA halved the downside. The distribution instead of an opinion.

Persi Diaconis shows a simulation can run for thousands of hours, look stable, and still be wrong. What he actually said about convergence and proof.

An AI swarm turned $1,000 into nearly a million? The number falls apart at its own source. Why consensus from correlated simulations is not an edge.