← Daybreak
BITCOIN, HONESTLY · A DATA-DRIVEN OVERVIEW

What 16 years of price history can — and can't — tell you

0 years·0 models

1 honest conclusion.

You cannot time Bitcoin's price. We prove it below. But you can read its cycle and forecast its risk. That's where the evidence-backed signal lives.

Built from ~30 models on Bitcoin's full daily history (2009 → mid-2026). We sell no predictions. The job is to separate what the data supports from wishful thinking.

THE WHOLE PICTURE

Sixteen years, four brutal crashes, one direction.

Bitcoin went from a fraction of a cent to over $120,000, a factor of more than 200 million. Every chart here uses a log scale. It's the only honest way to show percentage moves evenly across that range.

The crashes are savage and regular — −93% (2011), −85% (2015), −84% (2018), −77% (2022) — yet the long-run direction is unmistakably up.

overview

overview

loading overview
All-time high $126,219 (Oct 2025) · max drawdown −93% · today ~$62,000, about 50% below the peak.
PART 1 — THE TREND

Where the long run points.

Bitcoin's 16-year path follows a power law — a straight line on a log-log chart. Four independent methods agree on roughly where that trend lands by 2034.

When methods this different land within ~15% of each other, it's a real signal about the central tendency, not a price target. The realistic spread around it is enormous.

compare_all

compare_all

loading compare_all
History, an honest-width range of outcomes (P10–P90), and three model medians. Toggle log/linear top-right. Today ~$62,268.
FOUR METHODS CONVERGE — 2034 FAIR-VALUE TREND
$0kPower-law trend
$0kMonte Carlo (improved)
$0kGaussian Process
$0kBayesian power-law
rainbow

rainbow

loading rainbow
The popular version of the same idea: log-regression bands, each with its classic sentiment label. Hover any band. Today reads “HODL!”.
PART 2 — THE HARD TRUTH

You can't time the price.

We backtested every forecasting model the honest way: train on the past, predict the future it hadn't seen, roll forward, repeat 180 times. The benchmark to beat is the dumbest possible forecast, “tomorrow's price = today's price.”

Almost nothing beats it.

backtest_skill_heatmap

backtest_skill_heatmap

loading backtest_skill_heatmap
A sea of grey-slate cells. The more slate, the WORSE the model did than just using today's price; only a blue cell beats it. The best edge any model musters is under 3%, and it degrades from there. No reliable price skill, at any horizon, simple or AI.
hurst

hurst

loading hurst
Why? Bitcoin became a random walk as it matured. The memory measure sits right at 0.50 today, the signature of a memoryless coin-flip. Arbitrage traded the patterns away.
foundation_ensemble

foundation_ensemble

loading foundation_ensemble
Three frontier AI forecasters, same data: 2034 medians of $4k / $34k / $909k. The disagreement IS the finding — sophistication buys confidence, not accuracy.
PART 3 — WHAT YOU CAN READ

The cycle is real — and quietly fading.

Price level is unforecastable. Cycle structure is not. A wavelet analysis, which finds repeating rhythms, reveals a persistent multi-year cycle running across the whole history.

It's close to the famous “4-year halving cycle,” but slightly shorter (~3.5 years) and weakening as the market matures. Betting on a clean 4-year repeat means trading a signal that's going quiet.

wavelet_scalogram

wavelet_scalogram

loading wavelet_scalogram
A bright band at ~3.5 years runs the full history, but its strength fell roughly 19 → 6.5 → 4.6 across the three eras. Real, slightly shorter than folklore, fading.
seasonality

seasonality

loading seasonality
The “strong Q4, weak late summer” pattern, quantified. Strongest: October +13%, February +12%, May +10%. Weakest: August −8%, September.
valuation_overlays

valuation_overlays

loading valuation_overlays
Three classic cycle gauges at once. As of mid-2026: Mayer Multiple 0.81 (below trend), no Pi-cycle top, ~21% below the 2-year accumulation floor.
lppls

lppls

loading lppls
The bubble-detection model that nailed the Dec-2017 top to within days reads no bubble signature today — confidence ≈ 0%.
PART 3B — THE SCORECARD

Did the models call the cycles?

The cycle is real. The fair question is whether our models actually caught it. So we graded them the honest way. For each four-year-cycle turn, and using only the data available at the time, did each model call the turn in advance, lag it, or miss it?

The turns are the well-known ones: peaks in late 2013, 2017 and 2021, bottoms in early 2015, late 2018 and late 2022.

btc_cycle_scorecard

btc_cycle_scorecard

loading btc_cycle_scorecard
Almost nobody calls the turn on the day. The valuation and bubble models flagged the rich danger zone before every peak and the cheap zone before most bottoms — early, but never the exact date. Trend, regime and change-point models only confirmed each turn after it had begun.
forecast_baselines_expanded

forecast_baselines_expanded

loading forecast_baselines_expanded
More models, same test. We added a seasonal repeat, a vector autoregression on price and volume, a gradient-boosted up/down classifier, and Prophet. They land on or below the random walk too. Adding model families does not buy timing skill.
PART 4 — WHAT YOU CAN FORECAST

Risk is the one thing you can model.

The good news the backtest surfaced: volatility is forecastable even when price is not. A simple volatility model beats the naïve baseline by ~6.5%. Modest, but real and repeatable, unlike anything in price prediction.

And the risk that matters most — crash risk — is far worse than “normal” math implies.

garch

garch

loading garch
Bitcoin's “fear thermometer.” Volatility now ≈ 45%/yr — calm by its standards (long-run ≈ 105%/yr) — with clear spikes at every historical crisis.
regime_hmm

regime_hmm

loading regime_hmm
The market moves between three states: Calm grind (43% of days), Choppy (39%), Turbulent blow-offs (18%). Today reads “Calm grind / low-volatility uptrend.”
evt

evt

loading evt
Proper tail-risk modeling vs a naïve bell curve. A once-a-year worst day ≈ −32%; once a decade ≈ −74%. A bell curve would call −74% nearly impossible — understating crash risk by 2–4×.
PLAN-FOR-IT CRASH RISK — WORST SINGLE DAY
0%Once a year (fat-tailed model)
0%Once a decade (fat-tailed model)
0×Worse than a “bell curve” implies
0Crash/spike days per year a bell curve forbids
sim_jump_diffusion

sim_jump_diffusion

loading sim_jump_diffusion
Our most realistic simulation explicitly includes those ~7 jump days a year. Median path ≈ $334k by 2034 — but the downside tail is deep and real (P10 ≈ $40k).
PART 5 — PUTTING IT TOGETHER

The honest synthesis.

This is not advice to buy or sell. It's a map of what the data supports and — just as important — what it doesn't.

Long run

Trend points to mid-six-figures by the early 2030s if history holds. But the uncertainty is enormous and the timing is unknowable.

Short run

Unpredictable. Today's price is the best forecast of tomorrow's. Ignore confident price targets.

Cycle now

Below trend, in the cheaper accumulation part of the cycle, no bubble — by every classic gauge.

Risk

Elevated and fat-tailed. Plan for a −30% day in any year and a −70%+ day in any decade.

LIMITATIONS — READ THESE
  • One price series. No on-chain, derivatives, or order-book data.
  • The long-horizon forecasts can't be validated; with 16 years you can't backtest an 8-year prediction. Those cones are scenarios, not validated forecasts.
  • The trend assumes the past pattern continues. Power laws break; past growth doesn't guarantee future growth.
  • Forecast bands are imperfectly calibrated — too wide at short horizons, over-confident at long ones.
  • Descriptive ≠ predictive. These models describe what happened; they don't predict what's next.

Descriptive research, not financial advice. Bitcoin is volatile and can lose most of its value in a year — as it has four times in this dataset.

Appendix — all 14 charts & model inventory →