Trading research datasets — measured, not modelled

Which timeframe can each instrument actually trade?

Before testing a strategy, check whether the instrument can carry its costs on that timeframe. With stops expressed in multiples of ATR, cost stops being a property of the strategy and becomes a property of the market: round-trip cost ≈ 1 spread, risk per trade = stop ATR × ATR, and cost per R = spread ÷ (stop ATR × ATR). Measured on about 350,000 one-minute bars per instrument with each bar’s real spread (VT Markets, 2025-2026). Your broker’s spreads will differ — rerun the numbers with your own before trusting the last column.

On EURUSD the five-minute candle averages 2,7 pips and the spread is 1,1 — cost is 41 % of a typical candle’s range. There is no scalping in EURUSD, GBPUSD or oil at retail spreads. Gold is the friendliest instrument here, with ATR/spread of 22 even on M5.

How much did trading gold cost in each year?

The median bid/ask spread of XAUUSD, year by year, measured from tick data rather than quoted by a broker: every one-minute bar carries the spread it actually traded at. The series covers 2003 to October 2026 over about 7.9 million bars. It is not a smooth curve — the spread widened through the financial crisis, compressed through the 2010s and then jumped 53 % in a single year, from 0,38 $ in 2024 to 0,58 $ in 2025. At 0,68 $ in 2026 it is nearly three times the 0,24 $ floor of 2018. Anything that traded gold on the old spread is not trading the same market it was validated on.

Which of 392 strategy configurations survived their own costs?

392 configurations of strategy, stop type, stop distance and risk-reward ratio, each run on XAUUSD one-minute ticks with the broker’s costs already inside the numbers: 5 basis points of spread, 2 of slippage, 5 of commission — 9.5 in total. The table below is the top 20 ranked by Sharpe among the configurations with at least 100 trades. Ranking all 392 by Sharpe would put a different set on top, and not because those strategies are better. The CSV carries both rankings so the difference is visible.

What happens when you walk-forward a strategy with ten trades per window?

This is the least flattering table on the site. Eleven walk-forward windows over XAUUSD one-minute data, 2016-2026. The in-sample column ranks each configuration on the whole series — the number you get when you already know the future. The out-of-sample column is what the same configuration actually did on the following year. Nine of the eleven windows contain fewer than 15 trades, and a Sharpe over eight or ten trades is arithmetic, not performance: window 10 shows 708.12 with 10 trades and a profit factor of infinity. Two windows carry enough trades to mean something, and one of them lost money. We publish the table anyway, because the gap between the two columns is the lesson — and because a research shop that only publishes the pretty half is not a research shop.

† fewer than 15 trades — the Sharpe on that row is not a performance claim. 9 of 11 rows carry it; 2 do not.

Which strategies were verified, and which were discarded?

Thirteen strategies, each with the number that decided it. Three survive, four remain promising and six do not survive. In most of the failures the reason is not that the idea was wrong: it is that the costs ate it. A strategy with a good Sharpe on EURUSD at a 0,8 pip spread is a good strategy and a bad trade, and the register keeps the cost next to the metric for exactly that reason. The published metric column is not always a Sharpe — where the source paper published something else, it is kept verbatim.

How was this measured?