▲ Work in progress. This site is built and researched in the open: pages, numbers and findings may change or be wrong, and the record has gaps. Not advice. More
The Lab
Ideas traders use, put to the record
← the Lab · signals tested · write-up
#59 Purple whales: do big orders lead Bitcoin? No reliable edge
Past trades on public exchange files; no orders were placed. Every figure is a share, a correlation or a count.
Where the idea comes from
- Cumulative volume delta (CVD) adds up, trade by trade, the money spent by buyers who took a price minus the money from sellers who took one. A rising line means buyers were the more eager side.
- Split by order size, the same line is drawn for Bitcoin band by band: under $1k ("retail"), $1k to $10k, up to $100k to $1M, often coloured purple, and above. Traders call that band the "purple whales".
- The claim, common among traders: the purple band is the most reliable, and retail near the top of its range while purple sits near the bottom is a warning.
The question
- Can the chart be rebuilt from public trade records?
- If so, did the purple line lead Bitcoin's price over the next day, week or month, better than chance and better than the other sizes?
- Was "retail high, purple low" followed by falls?
Every rule and every bar was written down before any data was read.
How we tested
- Data: every Bitcoin trade on Binance's spot market from 2018 to 24 October 2026, and on Bitcoin's perpetual futures at Binance, Bybit and OKX (from 2020 or 2021), reduced to hourly sums per size band.
- The rebuild: each trade is placed in its size band, and the bands' buys minus sells are summed hour by hour. Each line is often shown on a 0 to 100 scale within a window (a week, a month, six months): 100 is the highest point in the window, 0 the lowest. We built four versions and kept the one that best matched 50 numbers read off recordings of the chart.
- The test: each day from mid-2018, each band's 0 to 100 reading, against Bitcoin's move over the next 1, 7 and 30 days. The early half of the years picks the best setting; the later half judges it.
- Five runs: Binance spot (the market the chart draws), the three futures markets one at a time, and all four added together.
The maths, step by step
- A rank correlation (Spearman's, called the IC here) asks: on days the signal ranked high, did the price also tend to rank high? It runs from −1 to +1; 0 means no link. Example: if the five days with the highest purple readings were the five best days that followed, it would be +1.
- Chance, measured. A link can appear by luck, especially when price moves overlap. So the signal is slid against the prices by a random 60 days or more, 1,000 times, keeping each line's own shape. The share of slid copies that show a link as strong as the real one is the p-value. Example: p = 0.20 means one slid copy in five did as well.
- The bar. We ran 414 checks in all. At p below 0.05, about 21 would pass by luck alone. So a purple result had to reach p below 0.0011 (0.05 split across the 45 purple checks that could count), in the later half, after leading in the early half.
What we found
- The rebuild works. 45 of the 50 chart readings matched ours within 25 points on the 0 to 100 scale, and the exact values read off still frames were about 6 points off on average. One trade counted as one order, on the 0 to 100 range scale, fitted best.
↑ US dollars
days from 2026-07-15 00:00 UTC →
— BTC
The last trade price at each day's start (00:00 UTC).
↑ 0 = the month's low, 100 = its high
days from 2026-07-15 00:00 UTC →
— $100-1k— $1k-10k— $10k-100k— $100k-1M (purple)— $1M-10M
Early August: retail near 100 and purple near 0, the warning pattern; read with the price above.
- Purple did not lead the price. Across the 414 checks, 20 passed p < 0.05 in the early half and 20 in the later half, about what luck gives (21). Purple's best early setting on each run, judged later:
| run | best early setting | early IC | later IC | later p |
|---|---|---|---|---|
| Binance spot | 6 months, next 30 days | +0.30 | −0.01 | 0.91 |
| Binance futures | 1 month, next 30 days | +0.10 | +0.21 | 0.049 |
| Bybit futures | 1 week, next 7 days | +0.05 | −0.00 | 0.99 |
| OKX futures | 6 months, next 30 days | +0.28 | +0.11 | 0.61 |
| all four | 1 month, next 30 days | +0.09 | +0.21 | 0.19 |
- On the spot market the chart draws, the early lead vanished. On Binance futures and the four together it held its sign, but well short of the bar.
- The warning. After days with retail at 80 or more and purple at 20 or less, Bitcoin on Binance spot was a little lower a day later in the later half (−0.45% after a weekly-window warning, p 0.06; −0.55% after a monthly one, p 0.04). Over 7 and 30 days, and on the futures markets, there was no such pattern. With 45 warning checks, a p of 0.04 is what luck gives.
What it means
- The chart can be rebuilt faithfully from public data.
- Over 2018 to 2026, the purple band did not lead Bitcoin's price better than chance or better than the other bands, on any of the five runs.
- Whether several weak signals together say more is a separate question. It would need its own rules, written down first.
Limits
- Our rebuild. Built from the known bands and windows, it matched 45 of the 50 readings, not all.
- Readings by ear and by eye. The 50 check points were read off recordings of the chart, and some recording times are unknown.
- A trade is not the same thing on each exchange. Binance merges an order's fills at one price; Bybit and OKX list each fill, which makes their orders look smaller.
- Daily readings. Someone reading the chart minute by minute may see things a once-a-day test does not.
Credits
- The order-size CVD idea is in wide use among traders; the bands and windows here are the commonly quoted ones.
- Data: Binance's public trade archive (data.binance.vision, CC BY-NC-SA 4.0), Bybit's public trade files and OKX's public trade records. Only hourly sums derived from them are kept.
- The shift test for overlapping returns is standard practice; the bars and settings are our own, written down before the data was read.
- PerpGremlin and Claude, Purple whales: do big orders lead Bitcoin?, TerminalGremlin, 2026: the tests, the code and the words.
← The volume clockTrend over months →
Educational research from public market data; any trading discussed is paper, never real. Nothing here is advice. The full notice.