Robinhood Chain / ETH research
Let AI modelsprove themselves
AI quantitative research for liquidity providers
Explore the process01 / Commit
Set the range
Then let time tell
Ranges trained on past observations
Historical tests and forward records stay separate
Waiting for verified data and model training
02 / Observe
One market
Different forecasts
Prices move forward, forecasts stay fixed
Every deviation remains open to review
Only observed pool candles can be revealed
03 / Evaluate
A useful forecast
survives the full account
Fees earned
Costs accounted for
Forecast accuracy and capital performance are separate measures
Every rebalance belongs in the ledger
LP return—
Buy & hold return—
Excess P&L—
View the accounting scope
One fixed range over the selected 24-hour window / No intraday rebalancing
Fees use real Swap and Flash activity under a marginal LP simulation / Gas units and slippage are assumptions
| Strategy | Net P&L | Return |
|---|---|---|
| Waiting for a completed calculation | ||
This simulation keeps historical prices and liquidity unchanged by the virtual position
It excludes MEV, market response and bridge costs / Swap step rounding can differ from contract execution
Historical simulations and forecast scores do not establish a trading advantage
Original historical rounds Reading the frozen batch
18 historical daily windows / Frozen for this historical batch
The full historical comparison is being calculated
Reading complete pool logs and the position ledgers
Fixed assumptions for the full batch / Separate from the selected-window controls
| Strategy | Total net P&L | Mean return |
|---|---|---|
| Waiting for a completed calculation | ||
| Benchmark | Mean excess / USDG | AI wins / Days |
|---|---|---|
| Waiting for a completed calculation | ||
This sample does not establish a repeatable AI advantage
Historical marginal simulation / Gas and slippage remain assumptions
Inspect every cost and round
Net P&L = inventory change + LP fees − gas − slippage − swap fees
Buy & hold begins with the same 50–50 assets and does not trade
| Strategy | Inventory | LP fees | Gas | Slippage | Swap fees | Total costs |
|---|---|---|---|---|---|---|
| Waiting for a completed calculation | ||||||
| Start date | AI | HV | EWMA | Hold | Coverage |
|---|---|---|---|---|---|
| Waiting for a completed calculation | |||||
AI mean excess after omitting each single day in turn
Waiting for paired comparisons
Does a higher execution cost change the result
Fixed slippage stress scenarios / 5, 10 and 25 bps
Excluded windows and reasons will be listed here
Each window retains its pool identity and complete source hash
Development strategy study Awaiting complete review
05 Oct 2025 → 03 Apr 2026 UTC / Arbitrum ETH / USDC / Known-history development
Development validation / Waiting for complete review
Reading the development study / Future independent results remain pending
Simulated capital 10,000 USDC per separate daily window / 5 bps primary slippage / All modeled costs / No compounding
| Strategy | Net P&L | Mean return | LP days |
|---|---|---|---|
| Waiting for complete reviewed development results | |||
| Benchmark | Mean excess / USDC | AI wins / Days |
|---|---|---|
| Waiting for complete reviewed development results | ||
All historical ETH periods are development evidence / Future independent results remain pending
Price features transfer from ETH/USDT to WETH/USDC / Quote basis risk remains
Why the original AI fell short
The original model optimized price excursions rather than net LP value
Its asymmetric ranges allocated an average 65–66% to ETH and increased entry swaps
Four falling days erased the excess earned on the other fourteen days
The new model learns inventory risk for eight LP widths from past price data
AI, HV and EWMA use the same width optimizer and allocate to LP every eligible day
Real chain fees and all modeled costs determine the final comparison
Separate model value from policy value
Replace only the conditional AI risk forecast with the static training risk curve
Keep the same fees, width optimizer, execution and dates to isolate the model contribution
| Strategy | Net P&L | Mean return |
|---|---|---|
| Waiting for complete reviewed development results | ||
Inspect development costs and rounds
| Strategy | Inventory | LP fees | Gas | Slippage | Swap fees | Total costs |
|---|---|---|---|---|---|---|
| Waiting for complete reviewed development results | ||||||
| UTC start | AI | HV | EWMA | Hold | AI action |
|---|---|---|---|---|---|
| Waiting for complete reviewed development results | |||||
Cost scenarios / AI mean excess against each benchmark in USDC
5 / 10 / 25 bps / Waiting for the fixed cost scenarios
Capacity exclusions and their observed liquidity shares will appear here
Separate immutable model and source hashes / Original 18 days are diagnostic evidence only
Every forecast
Leaves a record
01Commit before the outcomeA fixed record
Record the forecast timestamp, data cutoff and model version in advance
Once the outcome arrives, evaluate the forecast without rewriting it
02Go beyond hit ratesCoverage meets precision
A wider range is easier to cover, but may reduce capital efficiency
Assess quantile error, observed coverage and range width together
03Test against simple rulesEarn the complexity
Compare the AI model with historical volatility and EWMA baselines in the same market
A complex model should demonstrate the value of its complexity
04Account for the whole positionCosts included
Inventory value, fees, gas and execution costs all contribute to capital performance
Keep backtests, simulations and live results clearly separated
Research before capital
Build the evidence
Before the next allocation
Starting with one pool
Working toward a continuously evaluated LP research network