The Question
Start in January 2017 with $10 a day. Buy more when a single chart line says bitcoin is cheap, sell when it says rich, and by mid-2026 the wallet holds $164,000: built from just $5,200 of new money. The chart line is Power Law Position. The rest came from selling into strength and reinvesting the proceeds.
Plain daily buying with no selling, over the same window, costs $34,500 and ends at $239,000. A 13-indicator composite ends higher still, at $370,000, and because its sell proceeds fund later purchases, its lifetime out-of-pocket cost ($28,000) ends below plain DCA's (dollar-cost averaging: buying a fixed amount on a fixed schedule, whatever the price). Each route arrives at a different place. The question is what each is optimizing for.
Three strategies were tested against every meaningful entry point since 2017. The findings split cleanly: one wins on total dollars, one wins on capital efficiency, one wins on simplicity. None of them predicts price.
All figures below were computed with daily data through 12 June 2026, after the June 2026 weight unification (the pipeline, the dashboard, and the comparison tool now score days identically), and can be regenerated with ml/research/verify_articles.py. They move with bitcoin's price; earlier versions of this article cited a snapshot from a different date without saying so.
Three Strategies
All three start from the same base daily amount (e.g., $10/day):
Standard DCA. Invest the base amount every day, never sell. The baseline everything else is measured against.
Power Law DCA. One indicator: where bitcoin sits in its long-term logarithmic growth corridor (a fitted trendline, not a statistical law, as later pieces here establish). Buy 5x when deeply undervalued (grade A/B), 3x when leaning cheap (C+), 1x at fair value (C), sell gradually when overheated (D), sell aggressively at extremes (F). Cash from sales funds future buys.
Signal DCA. All 13 indicators weighted by SHAP importance (SHAP is an itemized receipt for a model's forecast: it scores how much each input moved the output). The composite score drives buy multipliers (5x/3x/1x/0.5x). When the composite drops below 3.8, selling begins. Cash from sales is redeployed during the next accumulation window.
What the Backtest Shows
Twelve start dates run from January 2017 to January 2025: the December 2017 top, the March 2020 crash, the November 2021 peak, the November 2022 FTX bottom. All backtests end at the data stamp above.
Each strategy wins at something different:
Signal DCA produces the largest total portfolio. It wins all 12 entry points on absolute value. From 2017: $370k against Standard's $239k and Power Law's $164k. The reason: it buys at 5x on more than half of all days, deploying more capital.
Power Law DCA is the most capital-efficient. It produces the best return per dollar out of pocket in 9 of 12 entry points (Signal DCA takes the other 3, all start dates in the 2017-2018 drawdown, the year-long slide from the December 2017 peak). From 2017: a $164k portfolio from only $5.2k of new money, a 31x return. The mechanism is the sell side. The D grade fires for about 11% of days, draining bitcoin into cash during the long climb to each peak. That cash funds the next accumulation phase. The market does most of the buying.
Standard DCA is the simplest. No decisions, no selling, no cash management. Signal DCA beats it on portfolio value at every entry point; Power Law beats it at 11 of 12 (plain buying from January 2017 edges out Power Law's smaller deployed stack on absolute dollars). It never costs more than the base amount on any single day. That is the trade.
Before reading those bigger numbers as a verdict on the signal, it is worth asking what produces them. The next two sections take that apart, because the answer is mostly not what the headline implies.
Why Power Law Is So Capital-Efficient
The verdict distributions diverge sharply:
- Power Law: 2% Strong Buy, 43% Accumulate, 44% Hold, 11% Reduce, 1% Sell
- Signal DCA: 55% Strong Buy, 31% Accumulate, 5% Hold, 3% Reduce, 6% Sell
(From a 2017 start. An earlier version of this article quoted a distribution computed by the data pipeline with an outdated weight set; these figures use the dashboard's weights.)
Power Law says Hold, invest the base amount, on 44% of days. Signal DCA says Strong Buy, invest 5x, on 55% of days. Signal demands much more cash during accumulation phases because it leans on the gas for most of the cycle; its sell-and-redeploy phases later return part of that outlay.
Power Law also sells earlier, and more often. Its D grade fires across the entire build-up to cycle peaks: $3.4k to $14.8k in 2017, $31.5k to $67.6k in 2021. Each tick converts bitcoin to cash. That cash then funds the next Strong Buy window. Sell, hold dollars, redeploy. Repeat with the cycle.
Where the Edge Comes From: Exposure, Not Timing
Signal DCA's $370k against Standard's $239k looks like a timing win. It is not one. The lead decomposes into three parts, and only the smallest of them has anything to do with reading the market.
Part one, capital deployed: nearly all of it. Signal DCA pulls $28k of pocket money plus its redeployed sale proceeds into bitcoin; Standard DCA pulls $34.5k but spreads it flat. Because Signal buys at 5x on 55% of days, it front-loads its buying into the cheaper, earlier years of a rising asset. More bitcoin, bought earlier, is worth more at the finish. On a fifteen-year uptrend that single fact explains most of the gap.
Part two, the sell-and-redeploy cycle: real but small, and mostly Power Law's. Selling into strength and rebuying into weakness recycles market profits into more coins. It is the engine of Power Law's capital efficiency, but the full composite rarely sells (6% of days), so it contributes little to Signal DCA specifically.
Part three, timing skill: zero, measured. A fixed-budget test isolates timing by forcing every strategy to invest the same total dollars, changing only the when. The result is +0.0% on average. Hold capital deployed constant and the signal's advantage vanishes entirely. It does not buy cheaper dollars than a calendar would; it buys more of them.
The verdict distribution is why this was always going to be the answer. The composite reads STRONG BUY or ACCUMULATE on 86% of all days in the 2017-onward backtest window, because the indicators flag accumulation opportunities and a structurally rising asset hands them a steady supply. A signal that says "buy, and buy harder" most of the time, applied to an asset that mostly went up, will win a DCA backtest by holding more, earlier. That is not a criticism of the signal. It is a precise description of what it does, and the reason the marketing-friendly framing ("the signal beat the market") is the wrong sentence. The honest sentence is: the signal sized the bet, and the bet paid because bitcoin rose.
The walk-forward weights test below closes the loop. If the specific SHAP-informed weights carried skill, they should beat naive alternatives. They do not: equal weights produce larger portfolios still, by being more aggressive more often. The weighting scheme is an aggressiveness dial, and the dial, not the cleverness, is what moves the result.
What SHAP Shows
XGBoost and LSTM (one a chain of yes/no questions about the data, the other a neural network that reads sequences in order), walk-forward validated from 2014 to 2026 (train on the past, test on a future the model never saw, slide the window forward, repeat), ranked Power Law Position first in 6 of 12 rounds, twice as often as any other feature, and placed it in the top five in 8, though mostly in small models (see the SHAP analysis for the full, corrected counts). SHAP points where the backtest points: one indicator carries most of the timing signal. The other twelve fill in the scenery.
Trend direction, on-chain activity, macro conditions: the rest of the panel adds context but does not improve the DCA outcome. Each still earns a chart on the dashboard because each tells its own story. Averaging them into a composite tends to mute the extremes rather than sharpen them. The signal is loudest when one indicator screams alone.
What If the Weights Couldn't See the Future?
The composite's weights were informed by ML runs over the same 2014-2026 history this backtest runs on. That is a look-ahead, and in June 2026 it was tested directly rather than just disclosed. A walk-forward variant re-derives the weights at each date using only SHAP rounds whose entire data window lies in the past: equal weights until mid-2018, then weights from the completed rounds, updated as each later round finishes. As a control, a third scheme uses flat equal weights throughout. All three run on identical mechanics (ml/research/walkforward_weights.py).
The result embarrasses the weights, in an instructive way. On absolute portfolio value, equal weights win 6 of 12 entry points and produce 33% more total value than the published weights; the walk-forward weights produce 35% less. On value per pocket dollar, the order reverses: walk-forward wins 10 of 12, published weights 2, equal weights none. No scheme shows timing skill. What the weighting actually controls is aggressiveness: schemes that read the market as cheap more often deploy more capital and end bigger; schemes that sell more often end smaller but far more efficient. The published weights are a middle setting on that dial, not a source of edge.
Two details are worth keeping. First, a practitioner deriving weights honestly in real time would have ended up weighting the 200-Week MA heaviest (34-54% across the epochs), not Power Law; the "Power Law is the anchor" framing is partly a product of hindsight. Second, this test isolates the weights only: the verdict boundaries and multipliers, optimized on 2014-2021 data, are held fixed in all three schemes, so the pre-2022 segments still carry their look-ahead. The conclusion that survives every variant is the same one the fixed-budget test gave: the system's outcomes are explained by how much capital it deploys and when it takes profit, not by which indicator gets 17% versus 8%.
What This Is Not
- No strategy times the market. Fixed-budget testing shows no timing alpha: no return earned by choosing the days, as opposed to choosing the amounts.
- Power Law sells early. Selling begins during the build-up to peaks, not at the top. Bitcoin sold at $30k will, later, trade at $60k.
- Signal DCA's large portfolio comes at a cost. It demands up to 5x the base amount on most days. The outperformance is not free.
- Bitcoin has been a structurally rising asset. All three strategies ride that current. Power Law models the curve directly.
- Only 3 cycle peaks exist in the record. The sell logic has been stress-tested against a small sample.
- The weights saw this history. The composite's weights were informed by ML rankings computed over the same 2014-2026 record the backtest runs on. The verdict boundaries were optimized on 2014-2021 and held out from 2022-2026, but the weights carry in-sample knowledge. The walk-forward test above measures what that is worth: the published weights add no timing edge over naive alternatives, and their effect is mostly to set how aggressively the strategy deploys capital.
- Recent entry points are underwater on efficiency. At the 12 June 2026 price, the 2024 and 2025 start dates return less than $1 of portfolio value per pocket dollar for both signal strategies. The strategies' record is built on entries that have had time to compound.
- No transaction costs or taxes are modeled.
- This is not financial advice.
The Bottom Line
For the simplest approach: Standard DCA. No decisions needed.
For the best return per dollar out of pocket: Power Law DCA. One indicator, clear rules, sells during overvaluation, redeploys market profits.
For the largest possible portfolio, with budget for 5x days: Signal DCA. 13 indicators, ML-weighted, aggressive accumulation.
None of them predict price. All three answer a different question: how much to invest, given current conditions. What separates them is how loudly they act on the answer, and whether they take chips off the table on the way up.
Try It Yourself
The interactive backtest lets you compare all three strategies with any start date and base amount.