Shuffle a market, watch luck build its own track record
Each permutation keeps every bar's size and shape but scrambles their order, so trends and patterns disappear while volatility and fat tails stay. Run a strategy on hundreds of shuffled copies: the histogram is what luck alone produces. The p-value is the share of shuffles that did at least as well as the real data. Switch to best of the whole search to see why a search's best result needs a much higher bar than one strategy. Only the first 70% of each dataset is used here, so the holdout is never touched.
Price path: real (dark) vs latest shuffles
Sharpe on shuffled data vs the real one
Catalog
Paper trading (no orders are ever sent)
loading…
Deploy a card that passed its holdout with oql paper deploy; check live against the backtest with oql paper reconcile.
Data collector
loading…
Ten quant mini-projects, each running live on the lab's data. Charts fill in as the work happens, and each project runs its built-in checks against a known answer before trusting the real result.
Live market (public exchange data, no account; updates every few seconds)
The efficiency ratio of the last hour is |net move| ÷ total distance travelled: near 1 the price went one way, near 0 it went back and forth. It is the quantity the efficiency research tries to forecast; here it is measured as it happens.
Efficiency ratio of a day (or hour) = |net move| ÷ total distance travelled. Trend rules are paid on efficient days and pay costs for nothing on choppy ones. Two questions, both on development data only (the first 70% of each dataset), and every run is counted in the ledger: can the next period's efficiency be forecast from what is known before it starts, and does switching a strategy on and off by efficiency help it. Each run checks itself on synthetic markets with a known answer first.
1 · Is efficiency predictable?
Walk-forward: each block is predicted by a model fitted only on earlier periods. Minute bars give hourly efficiency, hourly bars give daily efficiency. An opening-range breakout needed a forecast correlation of about 0.3 to be worth trading.
2 · Do efficiency gates help a strategy?
Every gate setting (trailing efficiency, or the forecast above; top or bottom third or half) is tried on one strategy, then: P&L by forecast third, and how strong a forecast would need to be. This is exploration. A gate is only tested for real when it is declared in a card's experiment, where the gate test reruns it on shuffled data.