Our Results: How We Test and What We Found

What these tests are

These are our own tests, run on past prices with small scripts. Each test applies one fixed set of rules to many overlapping windows of price history and records what the rules would have done in each. We built them to see how a grid and a doubling ladder behave, not to find a rule to sell.

Each figure below was read from a saved result file on 30 Oct 2026. The file name and the date of the file are shown next to it. The method for every figure on this site is on the method page. A figure we could not read from a file is not on this page.

How to read these

• A window. One start day and a fixed length: 30 days for the grid tests and 60 days for the ladder tests. A new window starts every day, so the windows overlap. 336 windows is not 336 independent tests, and the same stretch of price counts in many of them.
• Net. For the grid, profit from closed buy and sell pairs, minus fees, plus or minus the open gain or loss on units still held at the end, as a percent of the capital set aside for the grid. We read the percent as stated in the file.
• Median. The middle window when all windows are sorted. Half did better and half did worse. It hides the size of the bad ones, so the worst window is shown beside it.
• Lost. The share of windows that ended with a net below zero.
• Left the grid. The share of windows where price went below the bottom or above the top of the grid at some point. A grid that is left stops doing its job.
• Worst window. The single worst result in the set. It is one window, not a limit.

The rules that were tested

The grid. 20 lines spaced evenly across a range centred on the start price, with one unit bought at each line price falls through and sold at the next line up. The fee setting in the files is 0.001, which we read as 0.1 percent per side. The range is a half width of 5, 10 or 20 percent either side of the start price, as labelled in each table.

The doubling ladder. One unit is bought at the start. Each time price falls by a set percent below the last buy, double the last buy is added, up to 6 steps (63 units at full depth). Everything is sold when price is a set percent above the average cost. For the Bitcoin and Ether tests the fall is 3 percent and the take-profit is 2 percent. A window that has not hit the take-profit is judged at its end.

Summary of what was tested

Grid on Bitcoin, one year

Source file: btc_real_grid.json, file dated 17 Oct 2026. Data: public daily closes for Bitcoin from 2025-10-17 to 2026-10-16 (366 days). The price went from 108,168 to 82,166, a change of -24 percent over the period. Windows are 30 days, with 20 grid lines and a fee setting of 0.001.

What it shows: over this one year, in which price fell, the median window of every grid width ended at or below zero, and the narrowest grid was left by price in 95.2 percent of windows. What it does not show: how the same grid behaves in a rising or flat year, or with intraday swings. The file's own note says daily closes miss intraday swings, so the number of completed cycles is undercounted.

Grid on Bitcoin and Ether, by calendar period

Source file: multi_year_grid.json, file dated 17 Oct 2026. Data: public daily closes from an exchange market-data endpoint. Grid half width plus or minus 10 percent, 20 lines, 30-day windows, fee setting 0.001.

What it shows: the same grid gives different results in different years. The only positive medians are 0.08 percent for Bitcoin in 2024 and 0.08 percent for Ether in 2026, and both are small next to the worst windows in the same rows. What it does not show: why any period differs, or whether a later period will look like any of these.

Doubling ladder on Bitcoin, one year

Source file: btc_real_martingale.json, file dated 17 Oct 2026. Same Bitcoin daily closes as the first grid test (2025-10-17 to 2026-10-16). The fall that adds a step is 3 percent, the take-profit is 2 percent above average cost, there are 6 steps and the size doubles each step. Windows are 60 days.

What it shows: the ladder reached its small take-profit in most windows, and in 12.4 percent of them the price fell far enough to use every step. The worst largest open loss was 20.8 percent of the money put in. What it does not show: a ladder that hits its target often can still end with a large loss when it does not, which is the shape described on the martingale page. A high hit rate is not a measure of safety.

Doubling ladder on Bitcoin and Ether, by calendar period

Source file: crypto_ladder_periods.json, file dated 17 Oct 2026. Same ladder rules (3 percent fall per step, doubling, 6 steps, sell 2 percent above average cost), 60-day windows.

What it shows: the part of the table with a falling price, the 2026 rows, has the lowest share of windows reaching the take-profit, the highest share using all 6 steps and the deepest open losses. What it does not show: how a real account would have fared, because margin, liquidation rules and funding are not modelled here. See margin call for why that matters.

Grid on a gold-backed token

Source files: gold_proxy_grid.json and gold_proxy_periods.json, both dated 17 Oct 2026. The data is a token that is backed by physical gold, not spot gold and not a contract for difference, so spreads, swaps and weekend gaps are not those of a broker's gold product. The first file covers 2025-10-17 to 2026-10-16, where the price went from 4,439 to 4,164 (-6.2 percent). Windows are 30 days with 20 lines and a fee setting of 0.001.

The second file splits the token's history into three periods, using a grid of plus or minus 5 percent.

What it shows: the medians are small and close to zero, and the one falling period has the weakest row. What it does not show: how the rules would do on a real gold market, where the data and the costs differ.

The limits of every test on this page

• Hindsight. We chose the rules and ranges after seeing the data existed. We did not set them before an unseen test period, so these are not a clean out-of-sample test. See overfitting and forward testing.
• Costs. Only a fee per side is modelled. Spread, slippage, funding, swaps and taxes are not, and real costs would lower the results. See slippage and spread.
• Daily closes. Prices between closes are not seen, so fills inside a day and intraday losses are missed.
• Overlapping windows. The count of windows overstates how much independent evidence there is.
• Few markets and few periods. Two crypto assets and one gold-backed token, over about one to three years. That is a small slice of history.
• No live trading. No order was ever placed. A live bot faces rejected orders, outages, margin rules and your own choices.
• Fixed rules. A real user changes settings midway. These tests do not.

What we will not say about these results

We do not say a rule works, is safe or is worth using. We do not say a result will repeat. The tests exist to show shapes: where a grid is left, how often a ladder uses every step and how deep the open loss gets when it does. To check a rule of your own with your own numbers, use the grid calculator and the martingale calculator.