Methodology & Data Sources

Purchase schedule. Purchases land on UTC calendar days: daily, weekly (every 7 days), bi-weekly (every 14 days), or monthly (same day-of-month as your start date). In a month too short to have that day the buy lands on its last day, so a plan started on the 31st buys on the last day of February and again on 31 March. Exchange candles are UTC-aligned too. Doing the same math in local time shifts the whole schedule by a day for anyone west of Greenwich, which is why we don't.

Past and future. A purchase counts as made once its UTC date has arrived. If a plan's end date is in the future, the calculator runs it to that date and labels the headline figures as projected. The purchases still to come are priced at the live price the stack is valued at, or at the last known price when there is no live one. The chart marks today, and the purchase table lists the later rows marked Planned. Everything that describes what has already happened counts only the purchases dated today or earlier: the statistics under the headline (XIRR, max drawdown, best buy, fees paid), every panel on the Compare tab, the CSV export, the share card, the embed widget and the API. The stacking goal, FIRE and spend-down planners start from that same stack, the coins bought so far.

Fees. Each buy pays your percentage fee plus, if you enter one, a flat fee per buy: fee = amount × fee% + flat fee, and never more than the purchase itself. What is left buys bitcoin: btcBought = (amount − fee) / price. The full amount still counts as invested capital, so fees drag on returns here the same way they do on a real exchange. The fee figures we list for named platforms, and what each one leaves out, are on the exchange fee page.

Currencies. All math runs in USD. An amount you type in another currency (EUR, GBP, CAD, AUD, JPY) is converted to USD first, and results are converted back for display, both at one rate: the European Central Bank's latest reference rate, published once per working day and fetched through Frankfurter (api.frankfurter.dev). Past exchange rates are not used, so a result in euros is the dollar result at that one rate, not what each buy cost in euros on its day. The calculator prints the rate's date beside the result, and the CSV export carries the rate in a column of its own.

Today's rates are requested only when a currency other than US dollars is in use or the calculator's currency picker is opened. Until they arrive, a currency you have picked stays in US dollars, and the currency a shared link was made in is shown at once, at a fixed table dated 12 October 2026. If the rates have not arrived within about two and a half seconds, or cannot be fetched, the site uses the last ECB rates it holds or that table, and says in words that the converted figures are approximate. The Cost Basis Tracker converts each saved plan from the currency it was entered in at the same single rate, so the bitcoin, average cost and ROI of a plan entered in another currency are approximate too.

The default series is daily closing prices. From 20 July 2015 every day is priced at the close of Coinbase's BTC-USD daily candle for that UTC day. The calculator, the scenario pages, the studies and the API all use this series unless another is asked for. The current day is the exception: it has not closed, so its price is the latest trade at the time the series was fetched.

Before then, the prices are a daily average, not an exchange close. From 18 August 2010 to 19 July 2015 the series is a stored copy of blockchain.info's market-price chart, which blockchain.info describes as the average USD market price across major bitcoin exchanges. Each value is stamped 00:00 UTC, and where it can be checked against exchange candles it tracks the close of the day before its date. Measured on 12 October 2026 over the 4,101 days since July 2015 for which both exist, blockchain.info's value sat a median 0.045% from Coinbase's close of the previous day and 1.47% from its close of the same day. So a purchase before 20 July 2015 is priced about one day earlier in market time than one after it. The values are whole cents, which is coarse while bitcoin was under $1: at $0.06, one cent is a 17% move. It is a fixed copy, since that history no longer changes, and the same rows are used whichever exchange is selected. Nothing before 18 August 2010 is priced, so a plan cannot start earlier.

Those early rows are used as published, including two that look wrong. 15 and 16 September 2010 read $0.09 and $0.15, between days at $0.06. Coin Metrics' daily price for the same two days is $0.0604 and $0.0619. A second index disagreeing does not prove a trade did not happen, so the rows stay, and a purchase dated 16 September 2010 is priced at $0.15. Both sources were read on 12 October 2026: blockchain.info and Coin Metrics.

The Kraken series is mostly estimates. Kraken's public API returns only the 720 most recent candles of any size. Its daily candles therefore reach back about two years, and weekly candles are the finest that cover the years since 2015. We date each weekly close on the last day of its week, a Wednesday, and fill the six days in between along a straight line. Those days are estimates, not prices that traded. Totals come out close to the Coinbase series, but anything that reads single days does not: a fall inside one week is flattened, so max drawdown reads too small, and a best or worst buy can land on a price that never traded. Measured on 12 October 2026 with the default plan ($50 a week for 5 years), bitcoin stacked differed by 0.13% between the two series, while max drawdown read 47.1% on Kraken against 48.8% on Coinbase. For the same weekly plan run from 1 January to 30 June 2020, across the March 2020 crash, it read 27.3% on Kraken against 41.1% on Coinbase.

Every price says whether it was observed. Each day in a series is marked observed (a price the source reported for that day) or estimated (a straight-line value between two reported prices, or an earlier price carried forward to a day that has none). The marker travels with every purchase: it is in the CSV export's Price Kind column and in the API's responses, and each result counts how many of its purchases, and how many of the days checked for max drawdown, rest on an estimate. On the Coinbase series that count is normally zero. A file exported at a manual price has no Price Kind column, because none of its prices came from a market.

When an exchange cannot be reached. If the exchange you selected does not answer, the server first reuses the copy of that same series it fetched earlier the same UTC day. Failing that, it serves the other exchange's series, and if neither answers, the last series it holds, provided that is less than a day old. In every case the result carries whose prices they are and when they were fetched. The calculator and the scenario pages print a notice when the prices on screen are another exchange's or an earlier copy. A response with a hole in it is rejected, not filled in: more than 7 days between two Coinbase candles, more than 14 between two Kraken ones, or a series that stops more than 3 days ago.

Gaps use the last known price. If a scheduled purchase lands on a day with no price, the simulation carries the most recent earlier price forward rather than skipping the buy, and marks that purchase as estimated. A purchase dated after today is marked the same way, whichever price it is projected at.

One backtest is one start date. This comparison asks what the same plan did from all of the others.

What is run. Take your plan's frequency and its length, from its start date to its end date or to today if it ends later, rounded to whole months. The shortest is three months, and the plan has to be shorter than the price history. That plan is replayed from every possible start day in the history, one day apart. Each run buys the same fixed amount on its schedule and is valued at the price on its last day, so the result does not depend on the amount. Daily, weekly and bi-weekly runs buy every 1, 7 or 14 days. Monthly runs buy every 30.44 days, an average month, not on a calendar day-of-month. Your percentage fee is applied to every run. A flat fee, a starting stack and a yearly increase are not.

Which prices. Only the default series, whichever price source you have selected, and only up to the last finished day, which is normally yesterday: today has not closed. It can be limited to exchange closes alone (from 20 July 2015), which leaves out the years when the price is a cross-exchange average. If Coinbase cannot be reached, a copy of its closes fetched within the last day is used. With none in hand there is no comparison to show. It is never run on Kraken's estimates, and it is not shown at a manual price.

What is reported. The worst, 10th percentile, median, 90th percentile and best total return across the runs, the share of runs that ended below what was paid in, and where your own plan's return falls among them. With fewer than 20 runs only the worst and the best are shown.

Your own plan's return. The line marked “Your start” is your schedule measured the way the runs are: the purchases made up to your end date, or up to today if the plan ends later, valued at that last day's price in the series you have selected. It is not the headline return, which uses the live price and counts a starting stack. A plan with a flat fee or a yearly increase is not ranked, because the runs have neither.

Independent periods. The runs overlap almost completely: two five-year runs that start a day apart share 1,825 of their 1,826 days. So the number of runs says how finely the history was sampled, not how much evidence there is. The honest count is how many windows of that length fit end to end in the history, the days of history divided by the plan's length and rounded down. We report it next to the run count as independent periods. Checked on 12 October 2026, on prices through the day before: a five-year weekly plan had 4,073 runs and 3 independent periods (2 on exchange closes alone). A one-year weekly plan had 5,534 runs and 16 independent periods, and with a 0.5% fee 26% of those runs ended below what was paid in.

These are past outcomes. Every figure describes start dates that have already played out. None of it is a forecast or a probability.

Cost-basis percentile. The line on a share card that ranks your average cost uses the same replay, over all history and with no fee. It compares your average cost per bitcoin with the average cost of every run and reports the share of start dates that ended with a higher one, rounded down. Your percentage fee is taken back out of your own average cost first, which is why the line says “before fees”. It needs at least 30 runs, and it is left off for a plan with a flat fee, a starting stack, a yearly increase or a manual price.

The studies. Lump sum vs DCA, best day to buy and the start-date heatmap each carry their own method note on the page, next to the numbers it explains.