We sized Bitcoin positions by volatility. Did it help?

After testing ways to enter and exit Bitcoin, we tried a different question: can we improve returns by changing how much Bitcoin the account holds?

The idea is simple. Hold less when recent price swings are large. Allow more exposure when they are small. That can involve borrowing, so beating an ordinary Bitcoin holding is only part of the test. We also compared it with taking a fixed amount of extra risk.

How the sizing rule works

Once a day, the rule measures volatility from completed four-hour price changes. It divides a chosen annual volatility budget by that measurement to set its desired exposure. A 60% budget with measured volatility of 80%, for example, asks for 0.75 times account value in Bitcoin. The budget is a sizing input, not a promise about future risk.

Exposure cannot fall below 0.25 times account value or rise above 1.50 times. The account always stays long. With $10,000 of equity, 1.50 times exposure means about $15,000 of Bitcoin financed with $5,000 of borrowed cash, before costs.

We tested 40%, 60% and 80% budgets over 30 days, plus a 60% budget over 14 and 60 days. A sixth version capped exposure at 1.00 times, preventing borrowing. Fixed daily exposures of 0.75, 1.25 and 1.50 times, and Bitcoin bought once and held, completed the comparison.

The long-history result

The 60% budget over 60 days returned 74.70% annualized, compared with 68.36% for holding Bitcoin. Its average exposure was 1.05 times and its observed annualized volatility was 62.43%, versus 67.45% for Bitcoin. These averages describe what happened; they do not establish that the accounts had exactly equal risk.

That version beat Bitcoin under all twelve cost, delay and interest combinations in each of the four main periods: full history, 2019 onward, 2023 onward and the separate recent replay. It beat both fixed leveraged controls in all twelve cases since 2019, but only seven of twelve in full history and ten of twelve since 2023. These are encouraging comparisons, not probabilities that it will win in future.

The weaker results are part of the finding

In the reference price history, changing the volatility lookback from 60 to 14 days produced a much weaker result. The 14-day version lost to Bitcoin in every cost-and-delay-and-interest case since 2019. The 30-day version beat Bitcoin in ten of twelve cases over that period, but beat all three main controls in only eight. We cannot treat the best-looking lookback as validated after seeing these results.

The no-borrowing version lagged Bitcoin over the full history and from 2019 onward. Allowing leverage mattered. Even the promising 60-day version suffered a 78.14% observed drawdown in the main full-history test, reaching 78.40% under the other tested assumptions.

Separate annual accounts also show painful failures. The 60-day version lost 67.32% in 2022 versus Bitcoin’s 65.00%, lost 8.27% in 2025 versus 6.45%, and lost 23.21% through April 2026 versus 12.17%. It trailed Bitcoin in the rising years 2017, 2020 and 2021 as well. These annual accounts start afresh; they are not a decomposition of the continuous account.

Recently, extra exposure explains much of the gain

None of the volatility candidates beat Bitcoin and both fixed leveraged controls together in any recent sensitivity case. The higher recent return than holding Bitcoin therefore does not establish a useful forecasting signal. A fixed 1.50 times position did slightly better than the 60-day candidate.

A second price history weakened the result

We repeated the exact ten models using previously frozen Binance BTC/USDT prices, adding 3,360 source-and-scenario comparisons. This changes both the exchange and the quote currency. Each account is measured in its own currency: USD for the reference history and USDT for Binance. No exchange-rate conversion is modeled, so these are not equivalent dollar-investor results.

The 60-day rule still beat holding Bitcoin and both fixed leveraged controls under all twelve tested assumptions since 2019 on each source. But from 2023 onward, its advantage over all three controls fell from ten of twelve reference cases to zero of twelve Binance cases. At the main settings, Binance sizing returned +488.52%, versus +547.54% for fixed 1.50 times exposure and +367.89% for holding Bitcoin.

Starting in 2018 also exposed failures. On Binance, the 60-day rule lost to holding Bitcoin in both scenarios combining 0.30% trading costs with 20% borrowing interest. Across that period’s assumptions, its largest observed drawdown reached 82.61%. In the recent replay it returned +37.94%, versus +26.06% for holding Bitcoin and +38.14% for fixed 1.50 times exposure.

A follow-up data check found an important difference. Since 2019, the 60-day volatility input was unavailable at 716 of 2,674 Binance daily decisions, versus 60 of 2,674 reference decisions. The registered rule falls back to 1.00 times exposure when that input is missing. From 2023 onward, Binance had no such missing daily inputs, while the reference had sixty.

A small number of missing candles can invalidate a long volatility window for many days. These counts do not prove how much of the return difference came from the fallback. Price differences, quote currency and execution observations also changed. We then tested that missing-feature effect while keeping each source’s price values and execution assumptions fixed.

Download all 3,360 second-source results and availability counts, or read the registered second-source protocol. These are further checks on already explored market periods, not independent future results.

One missing candle changed two months of exposure

We completed that controlled test: 3,360 new cases with shared gaps in the volatility inputs, paired with the 3,360 existing original-input cases. Every trade price, execution observation, borrowing assumption and strategy parameter stayed fixed within each pair. Only the observations available to calculate volatility changed.

The reference history was missing the candle opening on October 25, 2025 at 16:00 UTC. Withholding that same observation only from Binance’s volatility calculation triggered sixty daily fallback decisions, from October 26 through December 24. Desired exposure became 1.00 times equity instead of roughly 1.22 to 1.50 times. Trade execution still used the original Binance prices.

That reduced exposure increased the Binance account’s terminal wealth by 13.41%, lifting its return from +488.52% to +567.44%. Fixed 1.50 times exposure stayed at +547.54%, so the ranking flipped. Across the twelve cost-and-delay-and-interest assumptions, wins against all three main controls rose from zero to ten. The reference account stayed unchanged because its gap was already present.

This identifies a material contribution from missing-data behavior in that historical comparison. It does not show that withholding data predicts returns. Shared gaps hurt other periods and models, and the 60-day rule still suffered drawdowns up to 82.88% across the main periods and assumptions.

The result makes complete, trustworthy input data a priority before changing the simulator’s strategy. We will not adopt those historical gap dates as a trading rule. Download all 3,360 new cases with their original references and decision-date explanation, or read the registered availability-test protocol.

The missing hours could not be recovered

We checked Coinbase’s public hourly records around all five missing four-hour candles, with acceptance rules written before retrieval. All three requests succeeded, but none of the twenty hours needed to fill those gaps was returned. We accepted zero repairs and kept the original histories unchanged.

Nine surrounding four-hour candles reconstructed from complete hourly records matched the frozen prices exactly. Three other existing candles had incomplete hourly coverage. The missing records do not tell us whether those intervals had no trades or whether the historical service lost or revised data.

This audit does not resolve the earlier strategy result. It shows why missing-data behavior needs its own test. No new performance scenarios were run on the empty supplement. Download the five-gap audit and neighboring-candle checks, or read the registered hourly-audit protocol.

Handling elapsed time weakened the advantage

We then tested a calculation that accounts for the time between actual price observations. A short gap can remain part of the risk estimate instead of clearing its entire history. This uses the change between known prices; it does not invent prices or reveal what happened inside the gap.

We kept the six original budget, lookback and exposure settings, plus two gap-limit checks for the 60-day version. The main rule permits intervals up to twenty-four hours and at most 5% extra time in the observation window. The additional checks use twelve and forty-eight hours. All definitions were fixed before the results: 2,688 new candidate accounts and 1,344 repeated fixed-control checks, with the old sizing results reused for comparison.

The comparison below uses the 60% budget, 60-day rule from January 2023 through April 2026. Returns are in each source’s native quote currency, with 10% borrowing interest, 0.06% trading costs and four-hour execution delay.

The reference account’s return fell from +559.34% to +481.69%, and its observed drawdown increased from 61.21% to 65.78%. It no longer beat fixed 1.50 times exposure. The Binance result stayed unchanged over this period because its relevant observations were already complete. The two price histories now give a similar verdict.

Since 2019, the 60-day rule’s wins against holding Bitcoin and both fixed leveraged controls fell from twelve of twelve tested assumptions to four of twelve on each source. All recent candidate returns stayed exactly unchanged. The twelve- and forty-eight-hour gap checks did not restore broad superiority.

Losses remain substantial. Across the main periods and assumptions, the central rule’s largest observed drawdown reached 78.40% on the reference history and 78.83% on Binance. Its reference return in 2025 worsened to −19.07%, versus −6.45% for holding Bitcoin. This is a more explicit treatment of missing data, but it did not produce a stronger trading result.

Download all 2,688 candidate results and 1,344 control checks, or read the registered elapsed-time protocol. Earlier tables above retain the original results; this follow-up changes their interpretation.

What the backtest can miss

The account checks each supplied candle’s low for a hypothetical 25% maintenance-equity breach. If triggered, it assumes a forced sale below that low, with extra costs. No forced sales occurred in these historical scenarios, but that does not establish liquidation safety: missing bars leave gaps, and a real lender’s margin rules, execution and rates can differ.

Reported drawdowns use observed four-hour closes; losses inside a candle can be worse. Missing volatility observations use the registered fallback, with no invented price bars. We preserved the earlier frozen price corrections. The long-history dataset ends in April; the recent replay uses a separately frozen input set. Neither is a live account record.

Definitions and complete results

• Download all 2,040 scenario results (JSON), including costs, interest, exposure and all failed variants.
• Read the registered sizing and account protocol.
• The previous spot buying-pressure experiments.
• Our earlier leverage study, which tested a different strategy and account model.

The elapsed-time study passed 578 software tests. All 1,344 fixed-exposure control curves and all 23,600 canonical reference decisions reproduced exactly. Future-price and price-unit checks also passed. Those checks support the implementation; they do not prove future profitability.