Why Simulate?
Bitcoin crossed $1,000 in late 2013 and brushed $68,000 in November 2021. Any strategy that bought more on the way up looks brilliant against that backdrop. Hindsight flatters the bold.
The backtest published here runs on that history. The trouble is that the next decade may not rhyme with the last one. A long bear, a flat stretch, a slow bleed to zero: none of these appear in the record, yet any one could define the next three years.
To probe that gap, three years of synthetic daily prices were generated across six scenarios and stitched onto real history (so indicators have a full warmup window). Standard DCA (dollar-cost averaging: buying a fixed amount on a fixed schedule, whatever the price), Power Law DCA, and Signal DCA were then run through each. The point is not prediction. It is to see which strategy holds up when the future refuses to cooperate.
A June 2026 revision replaced the simulation engine. The first version used Gaussian random walks (walks that draw each day's move from the tidy bell curve) with a single fixed seed, which contradicted this site's own tail analysis: bitcoin's daily returns have a Pareto tail and long-memory volatility, meaning the extreme days are not rare accidents but a standing feature, and a Gaussian path generator understates exactly the events that stress a DCA strategy. The current version draws Student-t innovations (a bell curve with fatter tails, so the simulated market produces the occasional violent day the way the real one does; degrees of freedom 3, matching the fitted tail index of 2.6-2.9) through a GARCH volatility process (calm days cluster with calm days, wild days with wild, instead of every day being an independent coin flip), and runs 40 independent paths per scenario. Tables report the median path with 10th-90th percentile ranges (the middle outcome, plus the band that contains the central 80% of runs: outcomes outside it happened, but rarely); charts show the median-outcome path. The engine is ml/research/simulate_forward_v2.py.
Limitation: on-chain indicators (MVRV, NUPL, Puell, Fees) and macro indicators (M2, DXY) cannot be simulated. Only price-based indicators contribute here, so real-world results would differ.
The Scenarios
Summary
What the Simulations Show
The signal strategies end bigger in every scenario's median path, but not through better timing. They end bigger by deploying more capital: Signal DCA pulls $52-55k from the saver's pocket over three years against Standard DCA's $11k. Per dollar deployed, the advantage vanishes. In the prolonged bear, Signal DCA and Standard DCA both finish at 96 cents of value per pocket dollar, to the cent. In the bull run, plain daily buying is actually slightly more efficient (2.07x against 2.02x). The strategies differ in how much they bet, not in how well.
The backtest finding holds in the forward test. The signal is a capital-allocation framework: it instructs the saver to invest more when conditions look favorable, amplifying the position in bull, bear, and sideways regimes alike.
The sell mechanism engages only when price climbs. Power Law DCA's pocket cost reveals it: in the bull run it deploys $35k (selling and redeploying through the climb), while in the bear scenarios it deploys the near-maximum $54k, the same as Signal DCA, because price never leaves the lower corridor and the strategy never sells. The sell-and-redeploy cycle, the engine of Power Law's capital efficiency in backtesting, needs a sustained climb. Only the bull and recovery paths deliver one.
The real risk: in a prolonged bear, Signal DCA pushes $55k of personal capital into a declining asset and ends with $53k (and as little as $34k on the 10th-percentile path). Standard DCA loses the same fraction of a much smaller bet. If BTC went to zero, Signal DCA loses five times as much real money. The signal does not protect capital. It deploys it.
Caveats
- Synthetic data, now with realistic tails. Paths use Student-t innovations (df 3) with GARCH volatility clustering, matching the measured tail index and volatility memory. Still missing: jump dynamics in the indicators that cannot be simulated, regime-switching correlations, and anything beyond a single-asset price process. Daily moves are capped at the historical worst (about 50%).
- Missing indicators. Six of 13 indicators (on-chain plus macro) cannot be simulated. The composite score uses only price-based inputs here.
- Forty paths, not one. Each scenario runs 40 independent seeds; tables report medians with p10-p90 ranges. The ranges are wide. In the hypervolatile scenario, Signal DCA's final value spans $39k to $145k across paths: path luck dominates within a scenario, which is itself a finding about how little precision a three-year projection can carry.
- Not a prediction. Which scenario arrives is unknowable. What is knowable is the trade: the signal does not call the bottom. It writes a bigger check at it. In a market that recovers, that check compounds. In one that doesn't, it bleeds.