How the market screen and setups work.

The screen is live in Perps Radar. For every Hyperliquid market it shows the tier, the tradability score and the facts behind them: 24h volume, open interest, impact spread, funding and how much the market moves. Presets are saved filters over the same facts. The screen describes markets only. It never says which way a market will go.

No long or short setups are shown. Version 1 of the rules below was measured on held-out history, after costs, against random entries in the same markets. No rule and timeframe passed all three tests of the gate at once: enough setups, a gain on average after costs, and a hit rate clearly above random entries. So none is shown.

The rules stay in research. They keep running on the live feed and every match is recorded, but none is shown. A rule is shown only once its own held-out record passes the gate. A new version of a rule starts its own record from zero.

A separate NYTSHIFT process watches every live Hyperliquid market, on the main DEX and the HIP-3 builder DEXes, through Hyperliquid's public websocket. It reads market contexts (funding, open interest, prices, 24h volume) and 15-minute and hourly candles. It never reads an account and holds no key.

Rules only look at closed bars. A bar that the feed missed is marked as a gap and no rule uses it until it is repaired. When the feed is stale the screen says so, and no old row is shown as current.

A market moves up a tier after 10 minutes above the line and down after 30 minutes below it, so a market near a line does not flicker.

Tradability is a score from 0 to 100 that answers one question: is this market practical to trade right now? It says nothing about direction.

• Liquidity, 40 points. 24h volume, open interest and impact spread.
• Movement, 30 points. Hourly ATR as a percent of price. Too quiet and too wild both score lower.
• Cost, 15 points. How much funding costs to hold for a typical stretch.
• Data, 15 points. A fresh feed and enough history behind the numbers.

Presets are saved filters:

• All markets. Every live Hyperliquid market, most practical to trade first.
• Liquid movers. Tier A and B, sorted by the 24h move measured in hourly ATRs.
• Volatile but tight. Hourly ATR in the top quarter of the screen, with an impact spread under 5 bps.
• Coiling. Bollinger bandwidth in its lowest 10% of the last 500 hourly bars (about 21 days).
• Funding extremes. Funding outside the market's own 5th to 95th percentile of the last 30 days.
• OI building. Open interest up 10% or more in 24h, with at least $1M of 24h volume.
• Waking up. Last hour's volume at least 4x the median of the 20 hours before.
• Stocks and indices. Tier A and B markets on Hyperliquid's builder DEXes (HIP-3).

When token reference data is switched on, the screen adds four columns from CoinMarketCap: the market cap band, the FDV band, Hyperliquid OI / market cap and FDV / market cap, with filters for both bands and for category. Mega $10B or more, Large $1B to $10B, Mid $100M to $1B, Small $10M to $100M, Micro under $10M, Unranked when CoinMarketCap reports none. The OI ratio divides Hyperliquid open interest in USD by the coin's market cap in USD, so a contract for 1,000 coins, such as kPEPE, does not change it. Where Radar's own category for a market differs from CoinMarketCap's, Radar's wins. A market whose coin has not been matched and reviewed shows nothing in these columns. Each column names CoinMarketCap and the age of its data, and none of it is used for orders.

These rules run in research, on tier A and B markets only. None is shown today (see where things stand). Indicators: ATR(14) and RSI(14) with Wilder smoothing, EMA(20/50/200), the 20-bar high and low before the current bar, Bollinger bands (20, 2.0), and the efficiency ratio over 20 bars. Normal volume is the median of the previous 20 bars. A rule needs its full lookback before it can match.

Every rule has a version. Changing any number makes a new version with its own record; old records are kept.

Not running yet. Rules based on funding and open interest (crowded longs, crowded shorts, squeeze breakouts) and on forced closes need at least 30 days of recorded open interest first. Hyperliquid keeps no open-interest history, so NYTSHIFT records its own.

What "inferred" means. Hyperliquid publishes no liquidation feed. A forced-close event would be inferred from open interest dropping fast during a sharp move on heavy volume, and it would always be labelled as inferred.

• Entry. The close of the bar that matched.
• Invalidation. The level where the rule is wrong, from the table above. R is the distance from entry to invalidation. If R is under 0.5 ATR the invalidation moves to 0.5 ATR; if it is over 2.5 ATR the setup is skipped as too extended.
• 1R and 2R. One and two times R from entry. They are measuring points for the record, not price targets.
• Score. 0 to 100 for how strong and clean the evidence was: trigger 30, context 25, market quality 25, data quality 20. It is not a confidence or a probability. If higher scores do not do better on history, the score gets a new version.

1. Entry is the open of the next bar, the first price anyone could get after seeing the setup. R is measured again from there.
2. The result is whichever comes first: the invalidation (-1R), +1R, +2R, or the close of the 24th bar.
3. If one bar touches both the invalidation and a measuring point and finer bars cannot tell which came first, it counts as the invalidation.
4. Costs come off every result: Hyperliquid's base taker fee of 4.5 bps on entry and on exit (scaled by the deployer's fee for HIP-3 markets), slippage of 1.5 bps a side for tier A and 4 bps for tier B, and funding paid or received for every hour held.
5. Results are in R, so leverage and position size never enter the record.

Against random entries. For every setup, 20 random bars are taken from the same market, month, timeframe and direction, with the same stop distance in ATR, and measured the same way. A rule that does no better than random is not shown, however good its own hit rate looks.

History is split in two. The first 60% is only used to check that the rules work as written. The last 40% is held out: it was never looked at while designing the rules. A rule and timeframe is shown only when its held-out record has:

• at least 100 setups,
• an average result above zero after all costs,
• and a hit rate whose 95% lower bound (Wilson) is above random entries.

Rules that do not pass stay in research and are not shown. Today no rule passes, so the screen is all that is shown. A shown rule is paused automatically once its live record reaches 50 setups and averages -0.15R or worse, or its hit rate's upper bound falls below random entries. Nothing is deleted.

Once a rule is shown, its history line always gives the number of setups and the random baseline. Fewer than 100 setups reads "Not enough history yet". The held-out record and the live record are never mixed.

There is no searching for parameters. If tuning is ever added, it will be walk-forward only, and this page will say how many variants were tried.