Our Thesis, Our Test, Our Limits
We believe Bitcoin is maturing into a digital store-of-value. At maturity, the BTC/Gold ratio should mean-revert around a stable band. This page lays out what we believe, how we test it, and where this model can fail.
The Thesis
Bitcoin: Digital Gold 2.0
Bitcoin represents the 5th evolution in monetary technology — after barter, precious metals, paper money, and fiat. Its fixed supply (21M cap), halving schedule, and decentralized nature make it a compelling store of value for the digital age.
Gold: 5000 Years of Trust
Gold has served as the ultimate store of value across civilizations. Its physical scarcity, durability, and universal recognition make it the benchmark against which new monetary technologies must prove themselves.
The Thesis: A Bet on Bitcoin's Maturation
The core thesis is not to time short-term moves. It is to bet that Bitcoin is maturing into a digital store-of-value and that, as it matures, its ratio against gold becomes mean-reverting around a stable band. While that maturation is in progress — measured in years, not weeks — where the ratio sits in that band signals which side of the cycle you are positioning into. If Bitcoin never matures the way the thesis claims, this model fails. The current halving cycle (2024–2028) is the first one in which the mean-reversion behavior can actually be tested out-of-sample.
Bitcoin Halving Cycles
Bitcoin's block reward is cut in half approximately every 4 years. These halvings create natural market cycles with distinct ratio characteristics.
Ratio Regime: Dynamic Rolling Window
The Ratio Regime indicator describes the BTC/Gold ratio's position using a Dynamic Rolling Window of the last 4 years (1460 days) with percentile-based bounds (2nd/98th) for outlier resistance.
- It asks: "Is Bitcoin historically cheap compared to the last 4 years?"
- This eliminates Look-Ahead Bias because no future data is used.
- A 30-day minimum hold period prevents rapid regime oscillation.
- Transaction costs (0.3%) are modeled into the backtested performance.
- The regime identifies the Primary Store: Bitcoin or Gold — designed for long-term holders.
Ratio crosses the 30th percentile of the 4-year window from above. Indicator: Bitcoin regime — held until the ratio crosses the 85th percentile.
Ratio crosses the 85th percentile from below. Indicator: Gold regime — held until the ratio crosses the 30th percentile.
Honest Assessment: How This Model Can Fail
Before the performance numbers below, you should know how this model can be wrong and where its claims are weakest. We surface these so you can decide for yourself whether the thesis is credible, not to bury them in a footer.
—Effective sample is n ≈ 3, not 4
We claim coverage of four Bitcoin halving cycles, but Cycle 1 (2012–2016) starts mid-cycle because BTC's reliable price history begins ~3 years in, and Cycle 4 (2024–2028) is partial. The effective complete-cycle sample is about three. Cycle-level statistical claims with n=3 are suggestive, not conclusive.
—Cycles 1–3 rode BTC's secular trend
Bitcoin appreciated roughly 10,000× from 2012 to 2024. Any strategy that was net-long BTC most of the time looks excellent in that window. The model's wins in those cycles are largely a function of capturing BTC's bull market, not of timing skill. The current cycle (2024–2028) is the first where mean-reversion behavior between BTC and gold can actually be tested independently of secular trend.
—Parameter selection on the same data
The percentile thresholds you see below are the in-sample winners of a parameter grid search performed on the full history (currently 30th/85th — these may shift as more data comes in and the grid is re-run). Reporting performance of a winning parameter on the data it was chosen from is a form of selection bias. The Significance Test section discloses the parameter sensitivity so you can judge how brittle the choice is.
—The model is regime-blind
It treats Bitcoin at $13 in 2013 the same way it treats Bitcoin at $100,000 in 2026. The macro context — interest rates, dollar liquidity, regulatory posture, institutional adoption — is wildly different across these regimes, but the model has no awareness of any of it. It sees only the rolling-window percentile of the ratio.
—Why BTC/Gold and not BTC/M2 or BTC/SPY?
The choice of gold as the denominator is defensible — both gold and BTC are reserve-asset candidates — but it is a choice. A reasonable skeptic can ask whether we picked the ratio that backtests best. We did not compare alternatives, and this page does not yet present that work.
—No live track record yet
Everything above is in-sample backtest analysis. Signal-change tracking is deployed in production, and the live track record will appear at the bottom of this page as regime transitions fire. As of today, zero out-of-sample transitions have been recorded. The honest answer to ‘how well does this model work in real time?' is: we don't know yet.
Stress-Period Breakdown
The strongest test of any regime-following model isn't the average case — it's the three Bitcoin bear markets where a buy-and-hold BTC position lost 75–80%. What did the model say going into each crash? All figures are computed from the same historical data used in the backtest above.
2018 Bear Market
Jan 7 – Dec 15, 2018
Bitcoin had surged 20× in 2017 to ~$16k, pushing the ratio to 12.46 — well above the cycle midpoint of 0.60 computed from the 2014–2018 rolling window. The model was in Gold regime from the moment the ratio cleared that midpoint, and remained there throughout the drawdown. Gold fell only 5% over the same period. The model held Gold through the entire crash.
COVID Crash (March 2020)
Feb 20 – Mar 23, 2020
The COVID crash was violent but brief — 32 days from BTC's pre-crash level to the bottom. The ratio was 5.92 against a cycle midpoint of 3.45, placing the model in Gold regime. The 30-day minimum hold enforced patience: even if a signal flip had triggered, the hold period would have prevented whipsaw action. BTC recovered fully within 6 weeks. Gold fell only 1%. A caveat: this is an unusually short drawdown and the model's 30-day hold is partly what made it a non-event.
2022 Bear Market
Nov 10, 2021 – Nov 9, 2022
Bitcoin peaked near $69k in November 2021 with the ratio at 34.98 — nearly 3× the cycle midpoint of 11.38 computed from the 2017–2021 rolling window. The model had been in Gold regime for months before this date and remained there throughout. As BTC fell 76% over 12 months, Gold fell 9%. By November 2022 the ratio had collapsed to 9.32, below the new midpoint — the model transitioned to Bitcoin regime, positioning for the recovery.
Operating boundary. The indicator was calibrated on three completed BTC halving cycles (effective n=3). It captures positioning relative to a 4-year mean-reversion band and is informative only while that band remains stable. When BTC enters a regime the band cannot describe — sustained one-direction trending, structural displacement of either denominator asset, or post-ETF institutional adoption curves outside the calibration window — the indicator should be read as silent, not as bearish or bullish.
Treat the regime as a positioning prior, not an execution strategy. The full-period rotation backtest at the in-sample winning 20/80 thresholds returns ~884% (CAGR 18.5%, Sharpe 0.64) versus a passive 50/50 BTC/Gold monthly rebalance at ~50,829% (CAGR 58.9%, Sharpe 1.21). The rotation thesis does not beat a static rebalance; what survives validation is the signal's ability to mark cycle position — which the runtime Model Confidence score (visible next to the gauge on the dashboard) is designed to monitor.
When the Model Confidence score is below 50, the ratio has drifted outside its calibration band on at least one sub-signal (trend pressure, band drift, whipsaw density, or volatility regime). Treat the indicator as silent in that condition.
Why This Indicator? Comparison to Alternatives
The BTC/Gold ratio is not the only macro-level indicator for Bitcoin positioning. A skeptic should ask: why use this ratio, and how does it compare to common alternatives?
Pi Cycle Top Indicator
Cycle top detectorStrength
Historically accurate at calling Bitcoin cycle peaks — the 111-day MA crossing the 350-day MA×2 has coincided with three major tops within days.
Limitation
Designed only to detect tops, not entrances. It offers no signal for when to re-enter Bitcoin after a crash, and has no Gold or store-of-value framing. It also has only fired 3-4 times historically — near-zero statistical sample.
vs. BTC/Gold Ratio
Complementary: Pi Cycle tells you about the peak. BTC/Gold tells you about macro regime across the full cycle.
Stock-to-Flow (S2F)
Supply scarcity modelStrength
Captures the structural scarcity argument for Bitcoin's long-term value. Widely cited. Reasonable for explaining the secular uptrend.
Limitation
S2F predicts absolute price levels, not relative asset positioning. It has significantly overshot on the 2020 and 2021 cycles — the model predicted $100k+ by end of 2021 and $288k by 2024. Price diverged badly from the model in 2022. It cannot tell you whether Gold is a better near-term store of value than BTC.
vs. BTC/Gold Ratio
BTC/Gold makes no price-level prediction. It is purely relative — which of two stores of value is currently favored by the ratio.
BTC/M2 Money Supply
Dollar-liquidity-adjusted ratioStrength
Strips out global dollar expansion to show BTC's value in real monetary terms. Useful for macro context on liquidity cycles.
Limitation
M2 data is monthly and often revised, introducing lag and noise. More importantly, it answers 'is BTC cheap vs. the money supply?' — not 'BTC vs. Gold.' The two questions are different: an investor might want to own the scarce asset with the best risk-adjusted performance relative to the other major scarcity asset, not relative to fiat.
vs. BTC/Gold Ratio
Roughly orthogonal. BTC/M2 is about fiat debasement; BTC/Gold is about which scarcity asset you hold.
Why BTC/Gold specifically?
Three reasons. First, both assets are monetary non-producers — they generate no earnings, pay no dividends, and derive value purely from scarcity and trust. Comparing them strips out noise from rates, earnings cycles, and fiat creation. Second, the ratio has 13+ years of daily data (Bitcoin's entire modern history), giving a meaningful backtestable series. Third, the mean-reversion thesis is falsifiable: if Bitcoin truly matures into digital gold, the ratio should stabilize around a cycle-relative band rather than trending unboundedly. That falsifiability is a feature, not a bug.
The honest caveat: we chose this ratio partly because it backtests well. That's selection bias and we name it explicitly in the Honest Assessment above. Other denominators (BTC/M2, BTC/SPY) are reasonable to ask about; we have not published a head-to-head with them, and we don't claim BTC/Gold is the best signal. We claim it's a legible, defensible one — both assets are monetary non-producers, both have a long history, the test is falsifiable, and the parameters are now locked.
The Decisive Test: Live Walk-Forward Benchmark
The hardest claim to verify on any quantitative model is "we are not still adjusting it." To make that auditable, the 30/85 thresholds were committed to our public repo in commit 409141a on 2026-05-22. From that date forward, every recorded snapshot is genuine out-of-sample data — the parameters cannot have been tuned to fit it. This is the one number on the site no in-sample optimization can touch, so we lead with it.
We replay what the deployed engine actually said since the lock — not a re-simulation — against a 50/50 monthly-rebalance comparator (same 0.3% cost per move), on two denominators: BTC/Gold and, more recently, BTC/SPY. The same framework agreeing across two independent denominators is far harder to engineer than any single backtest number. Here is the live record (also on the dashboard):
What you're looking at: a true out-of-sample test. The 30/85 entry/exit thresholds were locked at commit 409141a and have not changed since. Every dollar of return on the orange line was earned on data the model had never seen at the time the parameters were chosen.
Why this matters: backtests are easy to fit — you can tune any model to look brilliant in-sample. The walk-forward shadow is the honest version: parameters frozen, hourly snapshots, no edits. The gray dashed line is 50/50 BTC/Gold rebalanced monthly — a passive benchmark that requires zero forecasting skill. If the model can't beat it net of the same 0.3% transaction cost, the cycle signal isn't paying for itself. First meaningful read at day 90.
Why this is the headline: a model can always be tuned to look brilliant on the data it was fit to. It cannot be tuned to look good here — these curves are built only from signals recorded after the parameters were frozen. It is the slowest signal on the site, and the most truthful. Day 90 is the first meaningful read; before that the curve is collecting, not concluding. The in-sample check below is supporting context, not the proof.
Supporting Check: Beating Random Timing (In-Sample)
Secondary to the live record above, this checks the model against its own history. The right question is not "did it beat buy-and-hold Bitcoin?" — Bitcoin appreciated ~10,000× over the dataset, so its secular trend dominates any return test. The honest question: did the model's timing beat random timing between the same two assets, same trade count, same 30-day hold? We run 10,000 seeded random rotations and compare.
The punchline: measured over the period the model actually operates (after its 4-year band fills), it beats random timing with a p-value below 0.05 — random matched or beat it only ~1–2% of the time. We deliberately do not call that "statistically significant": the 30/85 threshold was chosen from a 36-cell grid on this same data, over ~3 cycles. A sub-0.05 number earned that way is encouraging, not proof.
Why we don't lean on it: a p-value computed on parameters chosen from the same data, over ~3 cycles, overstates the real edge — so even below 0.05 we treat it as suggestive, not settled. The decisive evidence is the live walk-forward record above, which no in-sample choice can touch.
Backtested Performance (In-Sample)
Important Disclaimer: These results represent historical performance of the Ratio Regime. Past performance as shown is not indicative of future results. See our full disclaimer.
Live Track Record
The only test that matters long-term: out-of-sample regime transitions detected and acted on in real time. This table records every signal change the model has produced since live monitoring began.