btc-cycle-lab

Changelog

Notable changes to btc-cycle-lab. Format follows Keep a Changelog; versioning follows Semantic Versioning.

One rule specific to this project: a change that alters what a number means is a breaking change, even when no interface moves. Adding a hypothesis to the scan changes every corrected p-value in it; changing an estimation window changes every abnormal return. Those belong under Changed with the effect on published results spelled out, never under Fixed as a detail.

[Unreleased]

Added

• The front page is one page, and it is a picture. It opens on where the price has been for two years with every buy and every sell the portfolio made marked on it, the level at which it sells next drawn as a line, and the calibrated 30-day interval standing at the right-hand edge. Nothing on that chart continues past today except the interval, which is labelled as one: the site’s own tests reject every rule that claims a direction, and a line running off to the right would be the part nobody reads the caveat under.
• “Which way next?” is answered with a chart instead of seven columns of p-values. Gold is each rule as written, grey is the same rule entered on dates picked out of a hat. One bar does clear its grey - halving +365d - and the page names it in the same breath as the reason it is thrown out, reading the p-value with the best trade removed straight from the result. The table itself moved to the evidence page.
• How far the price can travel, as a bar rather than a sentence. The 10- and 30-day intervals are drawn with today’s price marked inside them.
• A row of doors to the rest of the site, because the front page is now a summary and the workings had to stay one click away rather than one tab.

• The claims written before the outcome are counted on the front page. How many are recorded, how many have settled, and the date the next one settles. The ledger is the only page here that is not a backtest and it was the last tab; a reader could leave without learning the site writes its forecasts down in advance. The verdict sentence is forecast.ledger’s own, which refuses to give a score under twenty settled claims - a summary that wrote its own would be a summary flattering its own data.

• The front page says in plain words what the agent holds and when it sells. A reader who cannot read a p-value could previously leave the site without ever learning that a machine here trades a pretend portfolio in public - the one thing on the site with a plain answer. Four sentences now say what it holds, what it last did, the price at which it sells next, and the score against doing nothing. The same block carries the refusal in the same register: the rule behind those sentences could not beat entering on dates picked at random, and the block says so. test_publish.py fails if the numbers survive and the refusal is dropped.

Changed

• The site is set in the reader’s own system typeface. Everything was monospace at 14px - findings, tables, methodology and the licence footnote all at the same weight - so nothing said what to read first. It is now -apple-system, which resolves to SF Pro on a Mac or an iPhone and Segoe UI Variable on Windows 11, at 17px, with tabular numerals so columns still line up on the decimal point. No webfont is fetched: nothing in the pages sends a visitor’s address to a font host, which is a promise PRIVACY.md would otherwise have to make. Monospace survives inside <code> only. No published number changed; 0.93 in a column that carries three decimals is now printed 0.930.
• Table columns are named the way a reader would say them. The pages printed the DataFrame’s own column into the header, so a reader met p_worst_without_one and survives_correction with nothing to decode them by; the explanation now sits on the header itself. Booleans read yes and no rather than True and False, deliberately without a green/red palette: “rests on one trade: yes” and “survives correction: yes” want opposite colours, so weight carries it instead.
• Each section leads with its finding and folds its method away. The caveat under every result is still there in full, one click down, rather than printed at the weight of the result it qualifies.
• The intro plays once a week per visitor rather than on every load. Eight and a half seconds of skull, “PUMP!” and “-84%”, replayed at a reader who came back for the tables, on a site whose argument is that the fight is unfalsifiable. It is remembered in localStorage, with both the read and the write wrapped: a browser that blocks site data sees the intro rather than an error. The dashboard is exempt - there it is the whole screen and loops on purpose.
• The “Now” tab is called “The cycle”. Nothing distinguished “Today” from “Now” except clicking one, and the front page’s own link to that page had already been calling it the cycle.

Fixed

• The sell level is no longer drawn as a standing floor. trend_flip_level says in its own docstring that it is exact for one close and that quoting it as a standing line would be wrong: the day after, different days drop out of each moving-average window and it moves. The first version of the new front page drew it across two years of history labelled “sells below”, and the number went from $46,355 to $22,843 overnight while that was on screen. It is now a stub at the right-hand edge reading “if tomorrow closes below $X, it sells”, the card is labelled with the day, and the plain-words block says the price is worked out fresh each morning. No published number changed and the wrong version never left this machine; test_the_sell_level_is_never_quoted_as_a_standing_floor keeps it out.

• Tables no longer run off the side of a phone. All seven tables on the evidence page overflowed a 375px screen with nothing to say they had; a row is now a card carrying the column name it was read under.
• The cycle dial no longer draws over the chart beside it. Plotly measures its container when the figure is inserted, which in a two-column row is before the grid has settled, and responsive: true waits for a window resize that never comes. The pages now fit each chart to its container on load and whenever that container changes size.

[1.1.0] - 2026-09-07

Changed

• The published paper portfolio was restated: -14.5% since the last halving becomes -7.7%. run_paper applies the one-day execution lag itself and was being handed a series the backtest engine had already lagged, so every entry and exit landed a day late and the account paid for a day of moves it was not in. Every trade date on the page moves back one day (the 2024-08-11 sell is now 2024-08-10, and so on). The rule did not change and its verdict did not change: it still loses to holding, which returned +21.6% over the same window. The correction is printed on the page itself, not only here.

• The two constants that size the portfolio are now disclosed as fitted. The volatility target of 60% is BTC’s own median forecast volatility over this history (measured at 59.95%), not an outside convention; the 30% rebalance band was selected as the highest-Sharpe variant among those compared, on the same history the edge test then scores. The comparison is published in full, including the variant that scores higher than the one adopted. No number changed - what changed is that the page no longer presents either as though it came from outside the data.

• sizing now edge-tests the row it sizes with. It tested whichever variant had the best Sharpe, so the p-value on screen belonged to a strategy the portfolio does not run - vol target, EWMA band 10%, p = 0.40 - while the position came from vol target, band 30%, p = 0.079. Choosing the row on its Sharpe and then testing that same row on that same sample is the selection this project rejects elsewhere. The verdict is unchanged: neither is significant, and the sizing still earns no risk-adjusted edge. The best-scoring row is still printed by name.

• The rebalance band is no longer defended on its Sharpe. A sweep of all 61 values from 0% to 60% shows the adopted 30% ranks 34th, below the middle of its own grid; the best is 42% at 1.1886 against 1.1316. Neighbouring bands differ by 0.0149 on average against a standard error of 0.0202 for any one of these Sharpes, so the ranking is noise. From 52% the band is wider than the position ever moves: the rule stops trading and takes buy-and-hold’s Sharpe by construction, which is why the highest scores cluster at the right-hand edge. The band was not moved - adopting 42% for winning this sweep is the mistake the sweep exposes. It is now defended on turnover, which is a cost and does not move, rather than on a score that does.

• The volatility target is no longer defended on its Sharpe either, and for a different reason: that was never the right measurement. The target is a dial on how much risk to carry - across 39 values from 10% to 200% it moves the worst drawdown from -18% to -83% - and a Sharpe ratio is built to be indifferent to how large a position is. On that statistic the adopted 60% ranks 14th of 39 (1.1316 against 1.2151 for the best, at 30%). The number that does describe what the target is doing had never been published at all: at 60% the position sits at the no-borrowing cap on 50.2% of days, so on half of them the rule is not sizing anything, it is holding the asset. From a target of 160% it is pinned on 99% of days and the rule simply is buy-and-hold, its drawdown converging on buy-and-hold’s own -83%. The target was not moved; what changed is that it is defended as a risk choice rather than a measured optimum, with the half-of-days figure attached.

• Every sizing figure is restated: the cached volatility forecast had drifted from the prices it was built from. The stitch now prefers Binance from 2017-08-17, where Binance’s history begins, in place of Bitstamp; the cache predated that, so 3,305 of 4,035 daily forecasts behind the published sizing were computed from a price series the database no longer holds. The band’s rank in its own sweep goes from 34th of 61 to last, the target’s from 14th of 39 to 36th, and today’s position from 0.82 to 0.79. The estimator is unchanged and was verified deterministic: identical prices give identical forecasts, and appending four days changes nothing before the last day, so the whole move is the input. Both figures are on the page’s corrections log. The rank collapsing from mid-grid to bottom on a change of price source is the sharpest evidence yet for what the sweep already argued - that ranking these constants on this sample measures nothing.
• backtest/rank_history.py: a log of where each constant placed in its own sweep, one row per run, written every time the sizing is rebuilt and never pruned. The sweeps argue that ranking these constants measures nothing using statistics computed inside a single run; this makes the same argument by keeping the ranking and watching it move, which needs no standard error to believe. Rows are keyed on the day they were RECORDED rather than the last day of data used, because those come apart exactly when it matters - two runs over the same window can disagree once the window is restated underneath them, and both answers are worth keeping in the order they were believed. Charted on the agent page as each value’s place in its own grid, so a 61-value sweep and a 39-value one share an axis; below three distinct days the section says how many runs it holds and withholds the chart rather than drawing a trend through two points.
• The band’s turnover figures on the agent page are read from the sweep instead of written into the prose. They said “roughly 8.9x to 1.1x” and the refreshed data already says 8.8x to 1.2x - harmless the day it was written, wrong within a week of the daily job being switched on.
• The daily job now rebuilds the volatility forecast and the sizing tables. sizing_today.csv holds the position the agent page publishes, and it moved only when the command was run by hand - so the site could advertise a daily refresh while quoting a size computed on an arbitrary earlier day, with nothing on the page to show the difference. The step runs sizing --refresh (a bare sizing re-reads the cached forecast and would have been a no-op dressed as a fix), sits between ingest and paper because both paper and publish read what it writes, and is non-blocking: a failed refit leaves yesterday’s sizing in place and the day is still recorded. The four sizing tables are now committed with the pages that chart them.

Added

• backtest/sweep.py: the band sweep, with the statistics that disqualify its own argmax - the adopted band’s rank, the mean step between neighbours, Lo’s iid standard error (the optimistic one, stated as such), and the band from which the rule stops trading. Saved as sizing_sweep.csv and charted on the agent page with the frozen region shaded rather than cropped. target_sweep does the same for the volatility target and additionally records at_the_cap, the share of days the position is pinned at the no-borrowing limit - the column that says what a risk dial is doing, which no score can. Saved as sizing_target_sweep.csv and charted with Sharpe and drawdown on the same axes, because the comparison between those two lines is the point.
• publish/disclosure.py: an append-only corrections log rendered on the page whose figures moved, and the provenance note for the sizing constants. A restatement absent from that log is a restatement the reader never sees, so it is pinned by tests rather than by review.
• sizing_today.csv records target_annual_volatility, so the published provenance names the target the run actually used and cannot drift from it.

1.0.0 — 2026-09-05

First release. The project answers its question, and the answer is negative: on data from 2011 to 2026 there is no detectable halving effect in the price of BTC. Four independent checks agree, and each closes a different way of disputing it.

check	the objection it answers	result
control group	“everything rose, so what?”	difference vs NASDAQ p = 0.30
placebo category	“your method finds effects everywhere”	it does — raw p = 0.03 where nothing happened
walk-forward	“it worked in the past”	0 of 26 stable across 13 disjoint windows
specification curve	“you picked the wrong window”	160 specifications, none of it survives

The binding constraint is four halvings. No method fixes that, and none of the above is expected to change without a fifth.

Analysis

• Event study with inference across events, not days: t distribution with n−1 degrees of freedom, so n=4 gives a critical value of 3.18 rather than 1.96. A percentile bootstrap is reported alongside but never used to decide — at n<10 it produced ~30% false discoveries in testing.
• Hypothesis scan over 26 combinations of window, event category and macro phase, with Benjamini-Hochberg correction: 0 significant, raw or corrected. Replication out of sample is split by halving cycle, and one hypothesis passes that rule — event_credit_event_7d. It is reported with the two facts that sink it: its training-window effect was not significant (p = 0.42), so there was nothing to replicate, and its corrected q = 0.75. The rule was deliberately not tightened after seeing which hypothesis it would exclude; a significant_in_train column surfaces the fact instead.
• Walk-forward validation across 13 disjoint windows, with an embargo covering the target horizon.
• Control groups (NASDAQ, S&P 500, gold) compared paired by event, with a placebo run on the control itself to show the comparison can detect nothing when there is nothing.
• protocol_upgrade placebo category: the five consensus changes enforced on mainnet inside the price window. Complete by construction and pre-announced, so any effect found there measures the method. It produces the lowest raw p-value in walk-forward and the third-largest forecast coefficient.
• Specification curve: 160 combinations of price series, horizon, return type, abnormal-or-raw and estimation window, with inference by circular permutation of the whole curve rather than by counting significant specifications.
• Directional forecast (30-day horizon) with ridge-penalised logistic regression, scored against three baselines and thinned to non-overlapping labels. Verdict: no edge.
• Backtests against buy-and-hold with fees and slippage.

Data

• Prices stitched across Bitstamp, Coinbase and Binance with the seam reported and checked on overlaps (median divergence 0.06%).
• Macro from FRED as first releases (output_type=4), so available_from is when a number entered circulation. Measured median publication lag for M2SL: 43 days.
• Point-in-time contract throughout: every row carries both the observation date and the publication date, and all analysis filters on the latter.
• Hand-maintained event registry, every date checked against a named source.

Platforms

• Windows, macOS and Linux on Python 3.11 and 3.13, all six combinations verified on every push through the real installation path.
• run.py does the work — virtual environment, dependencies, dispatch — with btc.cmd and btc.sh as thin wrappers.
• Streamlit dashboard on port 8511.

Documentation

MIT licence, plus DISCLAIMER, PRIVACY, TERMS, DATA_SOURCES, SECURITY and CONTRIBUTING. Yahoo Finance prohibits redistributing its data, which matters for anyone hosting the dashboard publicly; only derived statistics are committed here.

---

Development history

Kept because several entries changed what the published numbers mean, and a reader comparing an old figure to a current one deserves to know why.

Methodological changes

• Specification curve (2026-09-04) — showed the apparent halving effect is drift: removing the pre-event baseline grows it eightfold, moving the baseline closer to the event nearly erases it, and it scales with the horizon without limit.
• Placebo category, cycle_extreme removed (2026-09-04) — the scan went from 23 hypotheses with 3 skipped to 26 with none skipped, so every corrected p-value from before that date is not comparable with one from after. A cycle extreme is identified from the price itself, so an event study around one recovers its own selection rule as a finding.
• Event registry sourced (2026-09-03) — two of twenty dates were wrong. mtgox_halt moved 2014-02-25 → 2014-02-24 and cme_futures 2017-12-17 → 2017-12-18, and a missing Mt. Gox event was added. Credit-event results changed; halving results did not, because no halving date was wrong.
• Inference switched from bootstrap to the t distribution — the bootstrap was miscalibrated at these sample sizes.
• Walk-forward replaced the single cycle split (2026-09-02) — one test window is one observation.

Bugs whose fixes changed results

• Event registry stored by merge instead of replace — the primary key is (name, date), so correcting a date left the old row behind and the event study averaged over both. The registry now replaces; prices keep their append-only upsert.
• Events with no estimation baseline entered as flat zeros — subtracting a NaN baseline made the whole row NaN, and nancumsum scored that as an event after which nothing happened. Affected permutation draws near the start of the history, which is the null distribution every specification-curve p-value is measured against.
• Benjamini-Hochberg poisoned by NaN p-values — the running minimum from the end contaminated the whole column.

Packaging and platform

• Launchers fixed (2026-09-05) — .cmd files had LF line endings, and cmd.exe re-reads a running batch file by byte offset, so call :label failed for labels far enough into the file. The POSIX launcher was renamed btc → btc.sh: Explorer hides known extensions, so the two appeared as identical entries and the unrunnable one sorted first. Double-clicking now either works or says why.
• English throughout and publication documents (2026-09-03), with a test that keeps it that way after {"kategoria": ...} reached a public repo. That test caught this very line, which is the behaviour intended.
• UTF-8 output on Windows — a legacy console code page turned an em dash in an event description into a crash.

This site is open source. Improve this page.