Bitcoin DCA worst case: how far down it has actually gone

Of 624 weekly Bitcoin start dates, 615 spent time below what they had paid in — 98.6% of them. The median plan was down 32.2% at its worst and 92 were down more than half. Yet the worst any start date has finished is 9.4% below cost.

Every argument for dollar-cost averaging is really a claim about the bad case: that spreading the buys makes the worst outcome survivable. That is testable. This measures how far down every weekly start date on record actually went, and how often.

The short version is that the journey has been much worse than the destination, and that the protection DCA offers is narrower than it is usually sold as.

How this is measured

Every Wednesday-to-Wednesday start date from Bitcoin's first daily close on 17 September 2014 to 26 August 2026 — 624 of them. Each buys $50 a week and never sells. For each one, three numbers: the deepest it ever sat below the money paid in, the worst fall in its balance from that balance's own peak, and where it finished. 4,362 daily closes, no projections. The same engine runs on the simulator, and the sources are in the knowledge base.

Every comparison below is drawn inside each plan's own window. Bitcoin's worst fall from an all-time high across the whole record is −83.4%, but that is a fact about the asset and not about any plan: a plan begun in 2020 was never exposed to 2018, and scoring it against a crash it slept through would flatter it for free.

Almost everyone goes underwater

615 of 624 start dates — 98.6% — spent time below what they had paid in. Only 9 never did, and they split into two kinds: 7 began between September 2015 and October 2016, close enough to a bottom that a decade of holding never once put them under, and 2 began in the last few weeks and have not been tested.

Being down is not the tail. It is the normal experience.

These are ranks within the 624 start dates on record, not odds. "Worst tenth" says where a plan sits among those 624; it is not a one-in-ten chance of anything, because the cohorts overlap heavily and there is only one price history underneath them.

The median is the number worth sitting with. Not the tail, not the crash of 2018 — the middle plan, the ordinary one, was down about a third of everything it had put in at some point.

The balance falls much further than the loss

These are two different questions and they get conflated constantly. How far below cost am I is the loss you could realize by selling. How far has my balance fallen from its peak is what the account looks like on the way down. A plan that is still buying can post a savage balance fall while never dropping far below cost, because each new purchase resets the basis.

The cushion is real, and it thins. Measured that way, the plan's balance fell less than the asset's in 622 of 624 cases, by a median of 7.7 percentage points. That figure is the middle of 624 gaps measured one plan at a time — deliberately not the distance between the two medians in the table above, which come from two separately ranked lists and answer a different question. But how much less depends almost entirely on how long the plan has been running.

Medians within each band. The cushion narrows at every step, and the bands are not doing the work: taken one start date at a time across all 624, the correlation between a plan's age and its cushion is −0.65.

The bottom band is where the mechanism shows. Contributions soften a fall only while they are large next to what you already hold. After a decade of buying, one week's $50 is a rounding error against the stack, and the plan falls like the asset it holds — for the plan that began 17 September 2014, −83.1% against the asset's −83.4% over that same window. The protection does not fail; it dilutes as the position grows — which means the cushion is thinnest exactly when the balance is largest.

The loss against cost behaves differently again, and the answer depends entirely on who you are being compared with. An asset has no loss against cost until you say when it was bought, so both alternatives below are priced inside each plan's own window.

Against the unluckiest buyer — one purchase, made at that window's worst possible moment — the plan sat less far below cost in 623 of 624 cases, by a median of 32.8 percentage points. That is the protection DCA is actually sold on, and it holds up.

Against the earliest buyer — one purchase, made on the plan's own start date — it is close to a wash. The plan was shallower in 299 of 624, and for the middle start date the two came out level. The spread is wide in both directions: in the worst tenth the plan sat 27.2 points deeper below cost than the single early buy, in the best tenth 15.2 points shallower. The reason is that spreading the buys raises your own cost basis as the price climbs, and your basis is the thing you are measured against.

The deepest plan on record shows all three at once, over one window: −56.5% for the plan that began 22 November 2017, −60.8% for a day-one purchase, −83.4% for a purchase at that window's top — which for a start date this early is the record fall itself, the one case where the whole-history number and the same-window number are the same number.

Anyone who bought it expecting a gentler-looking screen during a crash was sold the wrong thing.

Where they finished

The worst twentieth is the case people ask for, and it is oddly undramatic: −7.4%. The worst start date in the entire record finished −9.4% down. Set against a median of 2.07×, the distribution is lopsided — the downside has been shallow and the upside long.

That is a statement about a window in which Bitcoin rose a great deal, and it is worth being suspicious of. It is not a claim that the floor is −9.4%. It is the observation that across every start date so far, the terminal losses have been small while the mid-journey losses have been large, and those are the two numbers people mix up when they decide whether they can stand this.

The companion question — how long the bad stretch lasts — is measured separately in how long DCA buyers stay underwater.

What this does not tell you

Six things, plainly.

The cohorts overlap, heavily. Two start dates a week apart share every purchase but one, so 624 cohorts are nothing like 624 independent tries — they are a single price history read 624 ways. Every percentile above is a percentile of that one history, not a confidence interval, and it will understate how differently this could have gone.

The worst numbers are floors. 80 of these cohorts are still below cost, so their stories have not finished. Their worst marks can still deepen, and −56.5% is the worst seen so far rather than the worst this data will eventually show.

Survivorship is baked in. This measures buyers who kept buying through a −32.2% median drawdown against cost. It cannot measure the ones who stopped, and stopping feels most reasonable exactly when these numbers look worst. That is the gap between what the arithmetic says and what people actually do.

Past distributions are not forecasts. Every cohort here benefits from Bitcoin having risen over the measured window. A period where it does not would produce a different table, and nothing in this dataset rules that out.

Fees, spreads and taxes are ignored. A real $50 weekly buy carries costs this simulation does not model, and they fall hardest on the plans making the most purchases.

Down against what? Being 32.2% below your own cost is one fact; what the same money would have done elsewhere is another. The same schedule into other assets measures that side.

Run it on your own dates

The numbers above are one plan — $50 a week, every week, never selling. Change the amount, the frequency or the start date on the simulator and the whole distribution moves. It is free, and it runs on the same 4,362 daily closes this page does.