Skip to content
← Research

· Stopped

Can a dealer quote Robinhood Chain's stocks?

Answer

No. On single-stock/SPY pools, a dealer quoting Hyperliquid's HIP-3 price lost 10 to 18 bp an hour after each fill. The HIP-3 price is slower than the pool, so the dealer ends up filling the trades where the pool was right.

  • −10 to −18 bpthe dealer, an hour after each fill (s = 30 to 10 bp)
  • $16M/daynine stocks' v4 volume ($274M in 17 days)
  • 93%of it through pools with no hook; most fees 1–5 bp
  • −1.8 bppassive LPs, 5 minutes later (−3.2 bp an hour later)

How we measured

We read every Uniswap v4 swap on Robinhood Chain in pools holding one of nine stock tokens (2026-10-27 to 11-13) from the public RPC. Then we replayed a dealer that quotes the HIP-3 price ± s bp and hedges there. For each swap: would the dealer have quoted better than the pool, and what was that worth 5 minutes and an hour later? Nothing was deployed.

What else we found

  • On the nine stocks' pools, the dealer quotes better on 6% of the volume at s = 20 bp. In US market hours it loses (−9.6 bp an hour later). Overnight it wins (+11 bp). Weekends trade only $1.4M a day.
  • On the single-stock/SPY pools (AMD, MU, HIMS, TSM, INTC, PLTR; 17 pools, $29M in 7 days) it loses at every spread: −17.9 bp at s = 10, −15.3 at 20, −10.1 at 30. Only MU and weekends were positive.
  • Users were net sellers (NVDA −$0.84M over the 17 days). Stock tokens cannot be redeemed, so a dealer would pile them up.
  • Split by direction: swaps that move the pool toward the HL price cost LPs −7.8 bp. Swaps that move it away earn them +8.1 bp.

Why we stopped

We stopped the dealer.

If anything is worth keeping, it's the idea of quoting only off-hours, where it won. We haven't built that. We'll measure again if stock volume grows by an order of magnitude, or if we can read a price faster than HIP-3.

Limits

  • The reference is 1- and 5-minute candles, with a few bp of noise.
  • The dealer takes only the orders where it quotes better; routing and order splitting are not modelled.
  • v4 only; 17 days (7 for the SPY pools).

Measured with an AI (Claude). It wrote and ran the scripts, and every number here is a line of the committed output.