From Monday 9 November to Saturday night, BTC fell from $74,824 to $64,130 (−14.3%) and ETH fell 16.8%. Two explanations have been offered:

  1. Nvidia margin calls. Nvidia had a dramatic week. Polymarket's rules for one of its markets state that on 12 November NVIDIA announced Jensen Huang had resigned as CEO and from the board. The story goes that forced selling tied to Nvidia spread into crypto.
  2. Pre-IPO selling. Anthropic's IPO is close; Polymarket prices it for 19 November at 92%. The story goes that investors sold crypto to raise cash for it.

Both are plausible stories. Neither comes with public data on who sold what. Nobody publishes margin calls in real time, and nobody publishes the accounts that funded an IPO allocation. What is public is the tape: 1-minute prices for crypto and for Binance's 24/7 perps on NVDA and the rest of the stock market, Binance open interest and taker flow, Treasury and Nasdaq futures, and Polymarket's prediction markets on Nvidia's leadership and Anthropic's IPO, which work as timestamped news clocks.

So we tested what each story predicts. Here's what held up.

TL;DR

  • The stock market didn't dump with crypto. Over the same week the QQQ perp rose 2.4%, the SPY perp 1.7%, AVGO 14.7% and MU 12.1%. Only NVDA (−11.1%) and gold (−6.0%) fell with BTC.
  • Most of the drop happened when US markets were shut. BTC actually gained 0.9 log points during US regular hours. It lost 5.8 points on Saturday and 5.3 overnight.
  • Nvidia hit BTC twice, briefly. In the hour after each of NVDA's two big moves (−8% after hours on Wednesday, −6% pre-market on Thursday), BTC fell 0.6% and 1.35%. Those rank at the 6th and 2nd percentile of random hours in the same session, so it's a real effect, but it's about 2 of BTC's 15 points.
  • No sign of a margin-call cascade on Binance. Forced deleveraging closes positions. BTC open interest rose 3.5% over the week and was flat (−0.1%) through Saturday's −6.1%.
  • The Anthropic IPO doesn't show up in the tape. The three biggest jumps in Polymarket's IPO markets were followed by ordinary BTC hours (39th–85th percentile).
  • What did change was rates and the dollar: the 10-year yield went from 4.86% to 5.15% and the dollar index from 101.65 to 103.41, while gold fell with BTC. But minute by minute BTC didn't move with Treasury futures this week (ρ = 0.00), so that doesn't pin down the timing either.

Code and data: tick-microstructure-research (research/week46_crypto_dump.py). Everything comes from public data.

1. What fell, and what didn't

The week

Fig 1. Binance perps from Monday 00:00 UTC. Blue bands are US regular hours; grey is the weekend. Dashed lines mark the news moments tested below.

Mon 9 – Sat 14 Nov (UTC) week
BTC −14.3%
ETH −16.8%
NVDA perp −11.1%
gold (XAU perp) −6.0%
QQQ perp / SPY perp +2.4% / +1.7%
AVGO / MU / TSM +14.7% / +12.1% / +4.6%
AAPL / META / GOOGL / AMZN / MSFT +4.2% / +3.6% / +2.9% / +1.4% / −2.2%
crude (CL perp) −3.0%

If this were a broad deleveraging out of "AI risk", the chip stocks would have gone with NVDA. They didn't: Broadcom rose 14.7% and Micron 12.1%, and the Philadelphia semiconductor index closed Friday 1.1% above Monday (10,146 vs 10,034). The VIX ended the week lower (19.3 vs 19.9). NVDA had its own story, and crypto had its own week.

Where BTC's drop happened

By session

Fig 2. BTC's 15.4-point (log) week split by the US session each minute fell in.

US session BTC, log points share of the week's minutes
regular hours +0.9 23%
pre-market −3.2 19%
post-market −2.2 14%
overnight −5.3 28%
weekend −5.8 17%

By day, the losses grew through the week: −2.1% Tuesday, −3.5% Thursday, −4.4% Friday and −6.1% Saturday, on $18.8B of BTC perp volume, 4.3 times a normal Saturday. Whatever pushed crypto down did it mostly while US stock markets weren't trading.

2. Nvidia: two hits, both small

NVDA had two violent moves this week. We timed both to the minute, using the NVDA perp and Polymarket's Nvidia markets as clocks.

Nvidia news windows

Fig 3. Left: Wednesday after the close. Right: Thursday pre-market, when Polymarket's "no new NVIDIA CEO" market collapsed.

Wednesday, 17:00 ET. An hour after the close, the NVDA perp fell 3.2% in a single minute on $17M of volume, and was down 8.3% by 18:00. NVDA's own after-hours shares moved the same way (−8.3% on Yahoo's 1-minute bars). Polymarket's "Next CEO of NVIDIA" market saw its first trade 73 minutes later. Over the same two hours BTC fell 1.0% and the QQQ perp 0.8%.

Thursday, 08:50 ET. Polymarket's "No new NVIDIA CEO by 2028" fell from 0.27 to 0.08 within minutes, and the NVDA perp traded up to $79M a minute. NVDA fell 5.8% by 09:32, then reversed to +2.5% by 10:00. BTC didn't reverse: it fell 1.35% in the hour after 08:50.

Is a 1% BTC move after big Nvidia news unusual? We ranked BTC's move in the hour after each moment against 3,000 random hours of the same US session this month:

moment BTC next hour NVDA next hour percentile among random hours
Wed 17:00 ET, NVDA after-hours drop −0.60% −5.3% 6th (post-market)
Wed 17:00 ET, two hours −0.51% 18th
Thu 08:50 ET, "no new CEO" collapses −1.35% +1.1% 2nd (pre-market)

So the Nvidia news did reach crypto. On Thursday morning it looks like a real spillover: NVDA recovered within the hour and BTC kept falling. But the two hours together account for about 2 of BTC's 15 points.

Two more checks say the week's NVDA–BTC link was normal outside those moments:

NVDA vs BTC, 5-minute returns 21 Apr – 8 Nov this week
correlation, all hours 0.31 0.33
correlation, US regular hours 0.33 0.33
NVDA moving one minute before BTC (1-min returns) 0.022 (vs 0.31 in the same minute)

The one session where it rose is the weekend (0.28 → 0.62), and there NVDA barely moved (−0.7% on Saturday while BTC fell 6.1%). In our longer crypto-vs-stocks study, equity perps trailed BTC on weekends rather than leading it.

What would "margin calls" look like?

We couldn't find any public record of margin calls tied to Nvidia this week; brokers and lenders don't publish them. A forced-selling wave that reached crypto would still leave footprints on Binance:

  1. Open interest falls, because liquidated positions close.
  2. Taker flow turns one-sided, because forced sellers hit bids.
leg BTC BTC open interest (contracts) ETH NVDA perp QQQ perp
Tue 07:00–13:00 ET −1.5% +1.5% −3.2% −6.1% −1.1%
Thu 07:00–11:00 ET (Nvidia news) −1.7% +2.2% −0.7% +3.1% +1.2%
Thu 21:00 – Fri 02:00 ET −2.6% −0.2% −3.6% −2.0% −0.8%
Saturday (UTC) −6.1% −0.1% −8.3% −0.7% −0.8%
whole week −14.3% +3.5% −16.8% −11.1% +2.4%

Open interest went up during the Nvidia-linked legs and stayed flat during the two biggest ones. Positions changed hands; they weren't forced shut in bulk.

Taker selling was elevated but not extreme:

BTC perp, UTC day taker-sell share same weekday, previous 8 weeks volume vs usual
Thu 12 52.3% 51.2% ± 0.8 0.7×
Fri 13 52.7% 49.8% ± 1.2 1.4×
Sat 14 52.2% 51.1% ± 1.3 4.3×

Friday was the most one-sided day of the week, about 2.4 standard deviations above normal. Saturday's selling was huge in volume but ordinary in direction. Neither looks like a liquidation cascade.

3. The Anthropic IPO: priced in, but not visible in BTC

Polymarket's IPO markets moved a lot this week:

Polymarket market Mon 9 Nov Sun 15 Nov
Anthropic IPO in November 2026 0.79 0.97
Anthropic IPO on 19 November 0.20 0.92
Anthropic flips BTC (valuation above BTC's market cap) by 31 Dec 0.70 0.85

If investors were selling crypto in response to IPO news, BTC should fall right after these markets jump. It didn't:

moment jump BTC next hour percentile among random hours
Mon 15:34 ET "IPO on 19 Nov" 0.12 → 0.24 +0.34% 85th
Tue 06:40 ET "IPO on 19 Nov" 0.25 → 0.39 −0.13% 39th
Tue 07:30 ET "flips BTC" 0.72 → 0.56 +0.20% 78th

None of these is unusual. A slow, pre-planned sale to fund an allocation wouldn't need to react to news minutes, and we can't rule that out. But it would also have to explain why the selling hit crypto and not the Nasdaq, which rose. The "flips BTC" market rising from 0.70 to 0.85 is partly arithmetic: at $64,130, BTC's market cap is about $1.28 trillion, inside the $1.25–1.5T range Polymarket currently favours for Anthropic's IPO-day market cap.

4. What did change: rates and the dollar

daily close Mon 9 Nov Fri 13 Nov
US 10-year yield 4.86% 5.15%
dollar index (DXY) 101.65 103.41
VIX 19.9 19.3
SOX (chip index) 10,034 10,146

Rising real yields and a stronger dollar are the textbook reason for gold and BTC to fall together while equities hold up, and this week gold did fall 6%. But the textbook link isn't visible minute by minute. BTC's 5-minute correlation with 10-year Treasury futures (ZN) was −0.06, −0.12 and −0.18 in the three weeks before, and 0.00 this week. During Friday's overnight leg, when BTC fell 2.6%, ZN didn't move at all, and the Nasdaq-100 future fell only 0.8%. On Saturday, when BTC fell most, the futures markets were closed.

The macro backdrop fits the direction of the week. It doesn't fit the timing.

What we can and can't say

  • Nvidia: the news reached crypto, measurably, for about two hours. It doesn't explain a 14% week, and nothing in the open-interest data looks like forced deleveraging.
  • The Anthropic IPO: markets now price it for 19 November, but BTC didn't react to the news as it came in. A pre-planned sale of crypto can't be ruled out from price data alone, but it would have had to skip the Nasdaq entirely.
  • The rest: most of the drop came overnight and on a Saturday, on heavy volume, with open interest flat, no futures market open and no news market moving. We don't know what drove it, and the tape doesn't support anyone who claims to know for certain.

How FollowSM could protect a portfolio in a week like this

Nobody needed to know why crypto fell this week to be hurt by it, and nobody needs to know why to protect against it. A protection layer has to do three things: notice in time, act by rules instead of by mood, and keep working at 3 a.m. on a Saturday. Here is how this week's measured risks map onto what FollowSM provides.

1. Watch around the clock. BTC's net drop happened entirely outside US regular hours: −16.4 log points pre-market, post-market, overnight and on Saturday, against +0.9 during the session. FollowSM computes its signals 24/7, for crypto and for the 18 Binance TradFi perps. The NVDA and QQQ perps keep trading, and keep being scored, while the stock market is shut.

2. Watch the source, not only what you hold. On Thursday the shock started in NVDA, and BTC fell 1.35% in the following hour (2nd percentile). If you hold both, put them on the same feed and treat holdings that move together as one group: a flag on any member pauses the group.

3. React to events, don't forecast regimes. Across this series one trigger kept earning its place: per-event Binance taker sweeps. They warned of Polymarket repricing within a quarter of a second and were the only pause trigger that clearly beat randomly timed pauses, on BTC (Part 5). By contrast, gating hedges on VPIN did no better than random timing (Part 2). So the guard below reacts to sell sweeps and toxic-flow alerts and pauses new risk. It doesn't try to predict the week.

4. Fail closed, with hysteresis. If the check can't get fresh data, treat that as "don't add risk", and require a calm period before re-opening. Part 6 builds that properly, with tests.

A minimal portfolio guard on followsm-sdk, checked against the live API:

from datetime import datetime from zoneinfo import ZoneInfo from followsm_sdk import FollowSMClient NY = ZoneInfo("America/New_York") # Your holdings, grouped by what moves together: a flag on any member pauses the whole group. GROUPS = { "crypto": ("BTCUSDT", "ETHUSDT"), "ai_stocks": ("NVDAUSDT", "QQQUSDT"), } client = FollowSMClient() # free tier: fine for a few checks a minute; use a key for more def us_market_open(now: datetime) -> bool: """Regular hours only, Mon-Fri 09:30-16:00 New York time. NYSE holidays are not handled here.""" now = now.astimezone(NY) return now.weekday() < 5 and (9, 30) <= (now.hour, now.minute) < (16, 0) def portfolio_guard(now: datetime | None = None) -> dict[str, list[str]]: """Which groups to stop adding risk to right now, and why.""" now = now or datetime.now(NY) paused: dict[str, list[str]] = {} for group, symbols in GROUPS.items(): reasons = [] for sym in symbols: s = client.get_toxicity_snapshot(sym) if s.sweep_active and (s.last_sweep_side or 0) < 0: reasons.append(f"{sym}: sell sweep ${(s.last_sweep_notional_usd or 0) / 1e6:.1f}M") if s.is_toxic_alert: reasons.append(f"{sym}: toxic-flow alert (VPIN pctl {s.vpin_percentile})") if reasons: paused[group] = reasons if paused and not us_market_open(now): for reasons in paused.values(): reasons.append("US market closed: thin books, wider moves") return paused 

Wire the result into whatever places your orders: skip new buys and leverage increases for a paused group, and log the reasons. Running it on a schedule turns "I was asleep" into "the rule was awake".

What this does and doesn't do:

  • It stops you adding risk into a move. It doesn't sell for you. What limits the damage of a 14% week is position size and leverage, decided before the week starts. No signal replaces that.
  • It reacts after a move starts, not before. A sweep flag protects the minutes after the first prints, not the first fill. If you need to act inside the second, that takes the WebSocket streams, as Part 6 explains.
  • We haven't backtested this guard on this week. This article uses only public Binance, Yahoo and Polymarket data, not the FollowSM feed's own history, so we can't say how many points it would have saved. We'd rather tell you that than print a number we didn't measure.

Caveats

  • One week. Each test here is a handful of events ranked against random hours from the same month. A 2nd-percentile hour is unusual, not proof.
  • Minute bars. Anything inside a minute is invisible here.
  • No flow data. We see prices, volume, taker side and open interest on Binance. We can't see who sold, why, or on other venues.
  • Open interest for the TradFi perps is excluded. Binance's published series for those contracts has discontinuities this month (one-day jumps of more than 100%, values returning to the same level two weeks apart), so we only use BTC and ETH open interest, which update cleanly.
  • News content. We don't attribute moves to specific reports. Event times come from the prices themselves; the only fact we cite about Nvidia is the resignation, as stated in Polymarket's market rules.
  • Polymarket prices are probabilities set by traders, not facts. The IPO date and size are what the market expects as of 15 November.

Reproduce

git clone https://github.com/Follow-SM/tick-microstructure-research cd tick-microstructure-research && pip install -e ".[parquet]" matplotlib tmr crossasset fetch --end 2026-11-14 # 1-minute Binance perp klines python research/week46_crypto_dump.py fetch # Polymarket clocks, Binance open interest, Yahoo rates/futures python research/week46_crypto_dump.py analyze # research/results_week46.json and the three charts 

Yahoo serves 1-minute futures bars for only about 30 days, so the minute-level rates check can only be rebuilt from a copy fetched before mid-December.

Get started

FollowSM streams VPIN, ±1% book toxicity and imbalance percentiles and per-event taker sweeps for crypto and for 18 Binance TradFi perps, NVDA and QQQ among them, so you can watch a stock and BTC on the same feed.

👉 Get your API key at follow-sm.com/pricing

Not financial advice. Crypto, perpetual futures and equities carry substantial risk of loss.