DCA vs lump sum Bitcoin: when you buy, measured three ways

Deploying a lump sum at once beat feeding it in over twelve months in 68.3% of 571 Bitcoin start dates, a median 24.4% better — though the worst 5% finished 44.3% behind. Choosing weekly over daily moved the cost 0.02%, which is noise. Hesitating a year before starting lost in 77.5%.

Jon Phillips · 27 August 2026 · 10 min read

There are three versions of the question "when should I buy?", and they get answered as if they were one. They are not. One of them barely matters, one of them matters enormously, and the third is the one people actually get wrong.

All three are the same fact seen at different distances: in an asset that rises, money that goes in earlier buys more of it. What follows measures how much that is worth at each scale, using 467 start dates and 4,361 daily closes.

I build the simulator these figures come from, so weigh them accordingly. The prices are Yahoo Finance daily closes, 17 September 2014 to 26 August 2026, run through the same engine the site uses — so you can put the plans in yourself and get the same numbers back, or find that you cannot.

Fees are excluded throughout, and that cuts against one answer here harder than the others — 4,362 purchases against 144 is not a rounding difference on a real exchange. I come back to it, because it is the one thing that could overturn a section of this.

First: how often you buy is close to irrelevant

Take the site's usual plan — $50 a week — and hold the budget fixed while changing only the rhythm. Daily buys $7.14, weekly buys $50, monthly buys $217.41 — which come to within about a dollar a year of each other, because no whole number of cents divides a year three ways. That residue is why the score is what a bitcoin ended up costing rather than what the plan ended up worth: cost per coin normalises the last few cents of budget away, and comparing final values would not.

Across 467 start dates, buying weekly instead of daily moved the cost by a median of 0.02%. Weekly came out cheaper 47.8% of the time — a coin flip. The worst result in either direction was 0.36%.

That is the answer to the question most people are asking, and it is worth being blunt about it: at this scale the decision does not matter. Pick the rhythm you will actually keep to.

You do not have to take that on trust. Here are all three plans running side by side in the simulator — the same three amounts, the same start date, pinned to the same data this article was computed from. Over the full history they finish 4,362, 624 and 144 purchases apart and end up within about one percent of each other.

The monthly result is a trap

Monthly is a different story. Across the same cohorts it bought cheaper 85.4% of the time, a median of 1.43% below daily. That looks like a real finding, and it is not the one it appears to be.

Split those same cohorts by how much Bitcoin actually rose over each one, and the advantage turns out to be entirely a function of the rise. In the flattest group — the 51 cohorts where Bitcoin gained less than 100% — monthly bought cheaper only 37% of the time, and its median edge is +0.07%, which is nothing. In the steepest group it is cheaper 93% of the time. The gradient is the finding: the harder the market climbed, the more buying monthly paid.

Monthly deploys its whole month on day one, so its money is in the market earlier. That is what is being measured here — not frequency, but a few weeks of extra exposure, which is worth a great deal in a steep climb and nothing at all in a flat one. It is the same mechanism as the rest of this article at its smallest scale. Note what it does not license: read this as "buy monthly" and you have missed it. If monthly helped, it helped because the market rose — which you could not have known in advance.

Second: DCA against lump sum, when you already hold the cash

Now the version that matters. You are holding a pot of cash. Do you put it in, or feed it in over the next year?

This one has a clean structure that makes it easier to reason about than it first looks. Both choices end up holding bitcoin, valued at the same price on the same day. So the winner is whoever bought more bitcoin, and the holding period cannot change who that is — it only changes what the gap is worth. Everything is decided inside the deployment window.

Over twelve months, deploying at once bought more bitcoin in 68.3% of 571 start dates, with a median advantage of 24.4%. The longer the feed-in, the more it wins: over three years it is 79.9% and 111.0%. Over three months it is barely distinguishable from a coin toss at 57.5%.

Read that trend with one caution. A longer feed-in needs a longer runway, so the cohort count falls as the window grows — 610 start dates at three months, 571 at twelve, 467 at thirty-six. The later rows are measured over an older and shorter slice of history than the earlier ones, so part of the rise is which windows survive the filter rather than the feed-in length alone.

Nothing about that is specific to Bitcoin. Any market that rises on average punishes delay on average, and the mechanism is simply that money in earlier is money exposed longer. The same comparison has been run on stocks and bonds; I have not reproduced those studies and do not lean on them here.

Now read the other column. The worst 5% of those twelve-month cohorts finished 44.3% behind the spread-out version. One time in twenty, putting it all in at once cost nearly half the position.

That is the entire trade, and it is why "lump sum wins" is a bad sentence on its own. Deploying at once has the better average and the worse tail. Feeding it in is insurance, and like all insurance it costs something — here a median of 24.4% — in exchange for not being the person who put everything in the week before a drawdown. Which of those you want is not a question arithmetic can answer for you.

Third: the version people actually get wrong

Here is the one worth the most money, and it is not lump sum against DCA at all.

Someone hears about Bitcoin. They have $50 a week to spare, not a pot of cash. They decide it is not worth starting with a sum that small, so they will save for a year and buy something meaningful then.

Both arms set aside exactly the same money on exactly the same dates. What differs is when that money stops being cash and starts being bitcoin — which is the entire effect being measured, so it is worth being precise: the deposit schedule is identical, the purchase schedule is not.

Waiting a year lost in 77.5% of 467 start dates, at a median cost of 13.22%. Wait two years and the median cost is 35.81%. Someone who first heard about Bitcoin on 2 January 2017 and waited a year before starting ended up $87,807 behind — 41.1% — on identical contributions.

And the honest part: it won 22.5% of the time. Someone who heard about Bitcoin on 8 November 2021 — near the top of that cycle — and hesitated for a year came out 12.9% ahead, $3,026 better off. That is the highlighted row in the table above, and it is there because a table that only showed the losses would be an advertisement.

The p90 outcome is 13.3% in favour of waiting. So this is not a rule, it is a distribution with a clear centre of gravity: roughly three times in four, the year spent waiting for a better entry cost more than it saved.

What the three actually add up to

Frequency is noise. Deployment is a real trade with a real tail. Hesitation is the expensive one, and it is expensive in the least dramatic way possible — not through a bad decision, but through a reasonable-sounding delay.

On this record, a pot of cash deployed at once did better than the same pot fed in — usually, and with a tail that was brutal when it went the other way. Feeding it in bought a smaller worst case at a measurable median price. Waiting to build a pot before starting was the costliest of the three, and the least dramatic.

That is a description of what happened, not a recommendation about what you should do. I am not an adviser, this is arithmetic on historical prices, and the full disclaimer says so at more length. None of it says buy Bitcoin. It says that if you are going to, the calendar has cost people more than the rhythm did.

What this does not tell you

Fees are ignored, and they cut against the first finding rather than with it. The daily plan makes 4,362 purchases against monthly's 144. On a real exchange with a per-transaction minimum, that difference would dwarf the 0.02% the frequency section measures. So read it precisely: frequency is noise before costs. Add costs and it favours the less frequent plan — a different finding, and one this data cannot size.

Bitcoin has almost never had a flat stretch, let alone a falling one. The flattest group here is 51 cohorts out of 467, and "flat" means gaining less than 100% — on any other asset that would be a boom. The claim that monthly's edge depends on the climb is therefore tested against a thin and generous end of the range. A longer or worse history would test it properly, and this one cannot.

Every cohort here needed three years of runway. A start date counts only if three years of history follow it, so the most recent three contribute none — a filter, not a full sample. The deployment table is the exception, needing only each row's own feed-in length, which is why its cohort counts run higher.

These are overlapping windows, not independent trials. Cohorts start every week, so two consecutive three-year windows share all but seven days of their history. The 467 windows behind the frequency section span about twelve years — roughly four non-overlapping three-year periods. The win rates describe the record accurately; they do not carry the confidence a sample of 467 would normally imply. There are no error bars on them, and that is a choice rather than an impossibility — a moving-block bootstrap handles overlapping series like this one. They are absent because I did not compute them, not because they cannot exist.

Money not yet deployed earns nothing here. Cash waiting to be spent sits at zero interest throughout, which flatters the all-at-once arm in the last two sections — hardest where the pot sits idle longest. In the thirty-six-month row the fed-in arm holds roughly half the money back for years, so the widest gaps in that table are overstated.

Past distributions are not forecasts. Every measurement here comes from an asset that rose enormously over the period measured. In an asset that does not rise, every sign in this article flips — earlier money would be worse money.

The saving-up arm is simulated, not run by the engine. There is no "save it up first" mode yet, so that arm is rebuilt from the engine's own purchase ledger — same schedule, same prices, with only the timing of the spend changed. An earlier version of that rebuild stepped through the data by row instead of by date, silently lost a purchase, and produced numbers that looked entirely reasonable. The build now refuses to produce this page at all unless the rebuild reproduces the engine's own answer to the tenth decimal.

The simulator can reproduce one of these three, not all of them. The frequency comparison runs there in full, and the link above opens it. The other two do not: a plan's end date currently means both stop buying and stop valuing, so there is no way to ask it to feed money in for twelve months and then hold, and there is no save-then-buy mode at all. If those are the projections you came here for, it is worth keeping an eye on the tool.

One schedule is not the distribution. Everything here uses $50 a week into Bitcoin. How long buyers actually stayed underwater measures the other half of this question — not when you started relative to your own plan, but how long the plan spent below what you had paid into it.

Run your own

Every figure above comes from the same engine the simulator runs on, over daily closes back to 17 September 2014 and forward to 26 August 2026, with fees, spreads and taxes excluded. Open the simulator and put your own dates in — the lump-sum overlay is the one that answers the second question here directly.