MM Flow
Methodology

How the Live Picker works

One engine, six steps, closed loop. Markets are scanned, gated for quality, scored on after-cost EV, ranked by risk-adjusted score, tracked after the call — and the realized outcome is labelled back into the feature buckets so the next call uses what actually happened, not just what the rules said should happen.

  1. Step 1

    Market scan

    Every cycle, the picker fans across the full universe — canonical Hyperliquid perps plus every HIP-3 builder DEX. Each market is normalized to a single snapshot shape; downstream code never branches on which venue produced it.

  2. Step 2

    Quality gates

    Before any prediction, hard filters reject thin volume, low open interest, wide spreads, divergent oracles, and at-cap markets. The highest 'opportunity' raw score often comes from the worst market — the gates keep those out. HIP-3 markets pass stricter thresholds than canonical.

  3. Step 3

    EV after costs

    For every surviving market the engine estimates expected value on both sides (long and short), accounting for fees, half-spread, slippage, and funding over the expected hold. A market is only a candidate if its EV lower bound stays positive AFTER costs — not before.

  4. Step 4

    Rank by risk-adjusted score

    Candidates rank by EV / downside, not raw EV. A 5% expected gain with 20% tail loss ranks below a 1% gain with 1% tail loss. The picker surfaces the top N each cycle; anything below the score threshold drops to 'no trade.'

  5. Step 5

    Pick tracking

    Each call is remembered: entry price, current price, recent move, max favorable excursion, max adverse excursion. So a pick made at 12:00 is queryable hours later — was the call good? An event log records flagged / dropped / move-threshold-crossed events for every active pick.

  6. Step 6

    Label & learn

    When a pick exits the top-N, the engine records its feature signature at flag time alongside what actually happened — realized move, max favorable, max adverse, hold duration. These rows accumulate per feature-bucket (funding-z × premium-z × side). Once a bucket crosses the sample-size floor, the engine stops guessing EV from rules and starts reading it off the empirical distribution. The picker learns from its own calls.

Next

See the picker in action

Current picks, the no-trade state when nothing qualifies, and the event log tracking what happened after each call.

View Live Picks

The picker calls markets; it does not execute trades. Picks are informational. Past performance of past picks does not guarantee future outcomes. Every pick assumes round-trip taker fees, half-spread on entry, and funding over the expected hold — the EV shown is net of those costs.