← 블로그← All posts

2026-08-30

Our site showed a fake −52% drawdown for two weeks. Here’s the autopsy.

By FinScope Research · Data as of 2026-08-30

What our readers saw

From mid-August until August 30, the Monster Beverage (MNST) page on this site reported the stock 52.5% below its all-time high. The real figure was 6.5%.

Nothing about Monster’s business changed. What changed was arithmetic: Monster ran a 2-for-1 stock split, with split-adjusted trading from August 11, 2026 — and our price history ended up with the split applied to half of the series and not the other half. This article is the full post-mortem: how the corruption got in past three layers of validation, why our own defenses then made it worse, how we fixed it, and what we found when we audited all 815 tickers we track for the same disease.

What a split is supposed to do to a price chart

A 2-for-1 split halves the share price and doubles the share count. Nobody loses money. So every long-term chart you have ever seen silently rewrites history: after the split, all pre-split prices are divided by 2, so the series stays continuous. That rewritten series is the “adjusted close” — the number FinScope, and virtually every charting site, is built on.

That rewriting is exactly where things can break. For a few days after a split, a data provider is serving a freshly recomputed history. If the recomputation is incomplete, you get a mixed-scale series: some pre-split bars halved, others not. That is what happened here. After August 11 our upstream feed returned Monster’s pre-split history mostly unadjusted (around $95) with stray already-adjusted bars (around $47) scattered through it, plus the occasional empty bar.

How the corruption got past the guard

We assumed feeds break, so the pipeline has three write-time defenses. The relevant one here compares every fresh download against the history we already trust on disk — not price by price (a legitimate re-adjustment rescales every price), but day-to-day ratio by ratio, which survives any uniform rescale and snaps exactly at a scale seam. One 20%+ ratio break and the download is rejected.

Here is the subtle part. On the first post-split downloads, the feed’s unadjusted pre-split bars matched our on-disk history perfectly — because our history had been written before the split, it was on the same old scale. Every overlapping ratio agreed. The only novelty was a few new post-split bars at real (halved) prices, and new dates have nothing on disk to be compared against. The guard passed the download, and the file now contained a fake −52% cliff on the split date: $95-scale history welded directly onto $47-scale prices.

The paradox: our guard started defending the corruption

A few days later the provider began serving properly adjusted history — pre-split bars finally halved. The correct data had arrived.

Our guard rejected it.

Compared against the poisoned file, the corrected feed is the one with a scale seam. The comparison logic cannot know which side is right; it only knows the newcomer disagrees with the incumbent. So the pipeline did exactly what it was designed to do — kept the on-disk file — and thereby froze the fake cliff in place. The ticker stopped updating on August 19, our staleness alarm fired on August 26 after five days without fresh data, and the investigation that followed found the poisoned baseline.

The general lesson is worth stating plainly: a validation system that compares against history is only as good as the history it trusts. Once a bad write becomes the baseline, the same guard that should protect you will fight the correction. The failure window here was narrow — a mixed feed can only sneak through in the few days right after a split, while the old scale still matches — but narrow windows are precisely where data corruption lives.

The fix: repair the baseline, then judge each bar

The repair, shipped August 30, works in two steps against a known split (ticker, ratio, and split date are curated by hand — this never runs unsupervised):

1. Heal the baseline. If the trusted on-disk history still carries a price cliff at the split date, divide the pre-split bars by the split ratio. The check is idempotent: a clean file is untouched.
2. Judge every downloaded bar individually. Each fresh bar’s price is compared against the healed baseline for the same date. The ratio can legitimately be ~1 (correct), ~2 (bar still unadjusted — divide it), or ~0.5 (over-adjusted — multiply it). Anything else is substituted from trusted history, and if too many bars are unclassifiable the repair refuses to run at all.

The repaired series then goes through the same old ratio guard — the repair is advisory; the guard keeps the final word. 10,239 bars were rescaled. Once the provider fully heals, every ratio is ~1 and the whole mechanism becomes a no-op that can be retired.

The corrected MNST page now shows the split boundary as it actually traded: about $47.08 on August 6, $45.53 on August 11, no cliff. Monster’s famous all-time maximum drawdown — −96.0% into the December 1995 trough — is unchanged, because a uniform rescale inside one era cancels out of any drawdown ratio.

Then we audited all 815 tickers

One poisoned baseline invites an obvious question: are there others? We scanned the full history of every ticker we track for one-day price ratios of 1.8× or more in either direction — a gain above +80% or a drop beyond −44%, the signature a mis-applied split leaves behind — and cross-checked all 116 candidates against the provider’s current data and split records.

115 of 116 were real. The market simply produces moves that look like data errors:

No second poisoned baseline exists. The one flag the audit did raise was older and stranger.

The odd one out: JCI is secretly wearing Tyco’s history

Today’s Johnson Controls (JCI) is, legally, the former Tyco International — the 2016 merger kept Tyco’s corporate shell, so long price histories for “JCI” follow the Tyco lineage before 2016. And in July 2007 Tyco split itself three ways, spinning off Covidien and Tyco Electronics.

Here is the problem: in the source data that separation is not value-adjusted. The adjusted series steps down 53% on 2007-07-02, as if half the company’s value evaporated — when in fact shareholders received the spun-off stock. Every drawdown measured across that date is exaggerated. JCI’s displayed maximum drawdown is −92.2% with a 12.6-year recovery; measured only from after the separation, the 2008-crisis drawdown is −70.1%.

Interestingly, this is a legacy problem. We checked the four big modern spinoffs — GE spinning off GE HealthCare and GE Vernova, 3M spinning off Solventum, AT&T spinning off Warner Bros. Discovery — and in all four the provider credits the spun-off value through a fractional adjustment event (1.281:1, 1.253:1, 1.196:1, 1.324:1), leaving no artificial step at all. Old corporate actions predate that practice.

We chose not to rewrite JCI’s history — our series mirrors the source bar for bar, and inventing our own adjustments would trade a disclosed distortion for a hidden one. Instead the JCI page now carries a data note explaining the lineage and the 2007 step.

What you can take away as a reader

1. Distrust long-term charts for a few days after a split. That is the window where providers recompute history, and where mixed-scale errors live — on every site, not just this one.
2. When two sites disagree on a drawdown, check the adjustment first. Dividends, splits, and spinoffs are handled differently across providers; most “which site is wrong?” puzzles are really “which adjustment policy?” questions.
3. A −50% single day deserves a corporate-actions search before a panic. Real ones exist (see the table above) — but so do fake ones.
4. For long-lived tickers, ask whose history you are looking at. Mergers can quietly swap the lineage under a familiar symbol, as JCI’s Tyco past shows.

Data basis and limitations

All figures are computed from FinScope’s recorded adjusted daily closes. The MNST incident numbers (−52.5% shown vs −6.5% real, $99.94 vs $49.97 peak) compare the poisoned file we served against the corrected series now live. The audit covered 815 tickers’ full histories; the 1.8× single-day ratio threshold is a screen for split-scale errors specifically, not a census of all data issues. JCI’s −70.1% post-separation figure measures from the post-separation peak to the 2008-11-20 trough within the same series. Corporate-action semantics (spinoff crediting, lineage) follow the upstream source; where they distort results we disclose rather than rewrite.

How to view it in FinScope

The corrected MNST page and the annotated JCI page are live. For the concepts used throughout this post, see What is maximum drawdown?, What is recovery time?, and the methodology page.

Methodology: drawdowns are peak-to-trough declines from the running all-time high on adjusted daily closes. See the methodology page.