# Hyperquant BTC-relative strength — research and specification

Research reviewed 2 November 2026. Current website source specification: **BTC-regime-aware Balanced**, model version 9. This is a descriptive market screen, not a calibrated forecast or a claim of maximum predictive accuracy. It does not split all cryptocurrencies into two groups. Bullish/Bearish lists show the last confirmed directional assessment. The assessment uses relative price performance and momentum only. Data validity is tracked internally; historical membership is not current confirmation.

## Version 9: distinguish a stable BTC range from shocks and range breaks

The website uses `src/btc-structure.js` for current BTC context. The former `src/btc-context.js` implementation is retained only as a reproducible research comparator. A small net BTC return alone no longer defines sideways.

Require all 47 latest synchronized, valid, fully closed 15m OHLC candles. The volatility reference is the mean of 22 older true ranges, before the latest 24 returns, with a 2bp price floor. A new shock therefore cannot immediately inflate its own denominator. A prior 24-candle high/low box excludes the latest two candles and must first meet the range efficiency, slope, width and half-mean-drift checks; new highs/lows in an existing trend are continuation. Two closes beyond its boundary by 0.2 reference ATR classify a range break; one outside close is a transition. Recent mean true range above 1.5 times the reference or any of the last four true ranges above 3 times the reference classifies volatility expansion. This label does not assert that price is sideways.

A **stable range** requires no outside close or volatility expansion, 6h price-path efficiency ≤0.3, absolute least-squares slope ≤0.06 reference ATR per bar, total high/low width ≤6 reference ATR and mean-price drift between the two halves ≤1 reference ATR. A directional trend requires efficiency ≥0.35 and absolute slope ≥0.12 reference ATR per bar; remaining cases are transitions. These are explicit research thresholds, not universal definitions or tuned optimal parameters.

Within a stable BTC range, Bullish requires both 1h and 3h absolute coin returns, coin-minus-BTC returns and BTC-adjusted residual returns to be positive, plus a score ≥20 and higher highs **and** higher lows across adjacent completed 1h blocks by more than 0.1 coin ATR14. Bearish is symmetric. These block comparisons are current structure, not future-confirmed swing pivots. Beta is estimated on older returns excluding the latest 3h, shrunk toward 1 with 20 pseudo-observations, with raw beta bounded to [-2,3]. Near-zero BTC variance uses beta=1. The score blends tanh-normalized 1h/3h residual returns at 65%/35%, using older residual volatility with a 2bp floor. This residual is not a claim of independent alpha.

Outside a stable range the existing Balanced branch continues. The stable-range branch has no bootstrap probability; its additional strict-test p-values remain 1. A descriptive Bullish/Bearish assessment does **not** approve an entry. Current setup details show BTC regime, numeric strength, coin structure, distance from the recent mean and the assessment timestamp. They never attach current context to a historical setup. Missing OHLC cannot create a structure assessment.

An experimental entry policy separately requires stable BTC or an aligned BTC trend, aligned coin structure and an extension no greater than 2 ATR. It is **not wired into live Confirmed, entry eligibility or automated trading**. On 49 altcoins in 26 October–2 November, it retained 280 closed trades versus 3,490 original trades; net expectancy improved from approximately −$0.708 to −$0.211 per $1,000 notional after fees, base slippage and funding, but remained negative and did not beat V1 (+$0.039) across that week. Combining the new context with V1 produced +$0.230 on 167 closed trades in that week, but remained negative in earlier history, the last three days and stress costs. Earlier history for the primary filter was also negative. A seven-day block-bootstrap interval for improvement versus the original included zero. This is exploratory evidence for separating context from entry, not validated profitability. Full protocol, diagnostics, stress costs and rejected outcomes: `docs/backtests/btc-structure-v4-2026-11-02/report.html`.

The remainder of this document describes the unchanged Balanced branch unless stated otherwise.

## Research conclusions

1. **Separate rising and falling benchmark periods.** Morningstar's upside/downside capture framework supports asking different questions in these two regimes. Hyperquant uses conditional log-return *differences*, not capture ratios: dividing by small BTC returns would be unstable. This adaptation is our design, not a published intraday crypto trading rule. [Upside capture](https://admainnew.morningstar.com/directhelp/Glossary/Custom_Statistics/Upside_Capture_Return.htm), [downside capture](https://admainnew.morningstar.com/directhelp/Glossary/Custom_Statistics/Downside_Capture_Ratio.htm).
2. **Do not count correlated timeframes as independent confirmations.** Adjacent returns and nested timeframes share information. Block resampling preserves short-range dependence; it still relies on approximate stationarity and cannot guarantee uncertainty estimates during regime breaks. We use circular fixed-length blocks, not the stationary bootstrap's random-length blocks. [Politis and Romano](https://statistics.stanford.edu/technical-reports/stationary-bootstrap), [Shalizi, time-series methods](https://stat.cmu.edu/~cshalizi/uADA/16/lectures/26.pdf).
3. **Historical downside behavior is not automatically a predictive advantage.** Levi and Welch find plain beta can predict subsequent downside risk better than downside beta in equities. This is a counterexample to treating an intuitive conditional metric as a validated forecast; it is not a direct result for Hyperliquid. [RFS paper](https://academic.oup.com/rfs/article/33/6/2772/5571734).
4. **Crypto momentum evidence is mixed and horizon-dependent.** Liu, Tsyvinski and Wu identify a crypto momentum factor; Dobrynskaya finds momentum over weeks and reversal over longer horizons. These results do not validate 15-minute signals. Han, Kang and Ryu's October 2026 accepted manuscript reports much weaker cross-sectional profitability after realistic frictions and daily marking to market. We reviewed its abstract, not the paywalled full analysis. [NBER/JF](https://www.nber.org/papers/w25882), [momentum and reversal](https://doi.org/10.3905/jai.2023.1.189), [realistic-assumptions study](https://academic.oup.com/raps/advance-article/doi/10.1093/rapstu/raag011/8914819).
5. **A scan over many coins creates selection risk.** Holm adjustment controls family-wise error when individual p-values are valid, without requiring independent hypotheses. Here bootstrap p-values are approximate. In Balanced mode this stricter test is reported as an additional evidence tier; it is NOT a mandatory inclusion gate. Consequently Balanced does NOT claim a controlled universe-wide false-positive rate. [Holm original paper](https://www.ime.usp.br/~abe/lista/pdf4R8xPVzCnX.pdf), [Harvey, Liu and Zhu](https://www.nber.org/papers/w20592).
6. **Avoid optimizing thresholds against the current screen.** Backtest selection can overfit even with superficially reasonable holdouts. Operational changes requested during development are recorded below and must be frozen before a proper evaluation. [Bailey et al.](https://www.davidhbailey.com/dhbpapers/backtest-prob.pdf).

## Current model

Universe: active native Hyperliquid crypto perpetuals, excluding BTC. HIP-3 duplicates, stocks, commodities and indices do not enter this BTC-relative screen. All markets remain accessible under Markets. A low-activity prefilter of $100,000 rolling 24h notional volume avoids spending API budget on the least active markets; it is an operational choice, not a statistical theorem.

Use only closed, synchronized **15m** candles within the **last 12 hours**. No hourly or multi-day evidence enters the live score. Use up to 47 closes / 46 returns (11.5 hours of return intervals); the extra margin ensures that even the oldest candle start is less than 12 hours old at any point between updates. A hard wall-clock cutoff prevents older observations backfilling missing data. Require at least 32 aligned returns, 6 observations in each BTC regime and 90% timestamp coverage. Missing regimes remain insufficient instead of being inferred from older days.

BTC changes within ±1 bp are neutral. Never bridge missing candles or include forming bars (2-second close buffer). Each paired excess log return has weight `2^(-age / 3 hours)`: a return 3 hours older has half the influence. Weighted conditional means are calculated separately when BTC rises and falls, normalized by BTC return standard deviation, and transformed with tanh. Combine the original and outlier-trimmed normalized means equally before tanh, then shrink each regime component by `n/(n+4)`. Trimming removes the largest absolute 1% (at least one observation). The final RS is 100 times the average of the two shrunken components, bounded by −100 and +100.

### Balanced eligibility

For direction d (+1 Bullish, −1 Bearish):

- RS × d ≥4; each shrunken regime effect × d ≥0.01.
- At least 60% block-resampling sign support in **each** BTC regime, taking the weaker support across both block lengths.
- The outlier-trimmed mean must retain direction d in both regimes.
- Timing uses only the last **1h (65%) and 3h (35%)** relative returns. Normalize each by the last 3h relative-return standard deviation times sqrt(number of bars), with a 2-bp per-bar volatility floor, then apply tanh. A combined timing score ≥10 confirms Bullish; ≤−10 confirms Bearish. The 13 latest matched 15m closes must be consecutive. These overlapping horizons are one score, not independent confirmations.

A complete high-beta, defensive or mixed assessment clears the direction. Insufficient or stale data cannot create a direction, but preserves a previous assessment as explicitly dated historical context; it never preserves current qualification. There is no multi-day veto, half-window persistence gate or requirement for independent agreement across four timeframes. The 60% resampling support is **not** a win probability and is separate from JEV confidence; JEV thresholds are unchanged.

These windows, decay, thresholds and timing rules are explicit product hypotheses for a short-term screen. No cited source validates them as optimal or establishes forecast accuracy. Recent-heavy weights make the screen respond faster but reduce effective sample size. Counts are minimum availability safeguards, not claims of independent observations.

### Resampling and additional strict evidence

Use 2,047 deterministic paired circular-block replicates with two block lengths: ceil(cuberoot(n)) and twice that. Resample paired returns and their recency weights together. Blocks cannot bridge known internal timestamp gaps. Keep the wider 99% interval envelope and weaker sign support. The short, recency-weighted sample limits the interpretation of these approximate uncertainty estimates.

Approximate recentered one-sided p-values use a plus-one correction; directional p-values take the maximum across the two regimes. Holm adjustment includes both directions of all scanned universe members and treats unavailable evidence as p=1. This optional strict tier is not an inclusion gate. At this replicate count and universe size its resolution is limited; it is not a calibrated confidence claim. Repeated scans do not carry a controlled sequential error rate.

## Data delivery

All info reads pass through the same local server gateway. Metadata is cached for 5 minutes, perp DEX lists for 30 minutes, fees for 5 minutes, candle reads for 20 seconds and order books for 3 seconds. Equivalent in-flight requests are coalesced across tabs and scanner. Account balances and orders are not TTL-cached. Mainnet/testnet and wallet addresses have distinct keys.

Foreground chart/account requests have priority. The shared budget is 900 weighted units/minute, with background scans capped at 400; a 429 opens a shared cooldown of at least 65 seconds. Budgets are below Hyperliquid's documented 1,200/IP/minute but cannot account for unrelated applications sharing the IP. Chart WebSockets provide live updates; REST reconciliation runs about every 5 minutes while healthy and every 30 seconds while recovering. The scanner fetches only 48 recent 15m candle intervals per market, persists them locally, then streams updates. Scoring updates on each 15m close. It does not fetch hourly or multi-day history, or separate 1h/3h/12h candles. [Official rate limits](https://hyperliquid.gitbook.io/hyperliquid-docs/for-developers/api/rate-limits-and-user-limits).

Statistical recomputation checks for new closed candles every 3 seconds and reuses unchanged results. The RS scanner does not request or subscribe to order books. It scans active native markets without a minimum-volume gate; volume only prioritizes scan order and does not affect the score or directional classification. Startup restores evaluations from cached candles before network reconciliation; changing the selected chart market does not restart the scanner. Metadata refresh runs independently every 5 minutes. Initial scanning and background updates have distinct empty-state messages. [Official WebSocket subscriptions](https://hyperliquid.gitbook.io/hyperliquid-docs/for-developers/api/websocket/subscriptions).

## Direction and data-validity architecture

State schema 2 separates two concerns; model version 9 invalidates cached earlier assessments:

1. **Closed-candle assessment:** evaluate the existing two-regime BTC-relative performance and 1h/3h momentum agreement. Publish `directionState` with direction (bullish, bearish or null), candle close time, score and classification. A complete neutral/mixed assessment clears a previous direction; a complete opposite assessment reverses it immediately. Incomplete evidence is unknown, not a neutral verdict.
2. **Data validity:** internally track current/updating/stale/offline evidence. A saved assessment never supplies a current signal when the latest candle pair is missing. Membership and ranking use the last confirmed direction and score, with an alphabetical tie-break, so missing data does not make rows disappear. Complete new assessments can still change the list.

The server persists dated assessments alongside candles and restores them for the matching model/state schema. Partial startup snapshots retain historical rows; confirmed delisting removes them. A successful new neutral assessment supersedes saved history. The client retains the previous snapshot on transport errors and suspends current confirmation. There is no artificial holding period, extra confirmation delay or newly tuned score threshold.

The list has no data, timestamp, qualification or details footer. The row tooltip and accessible label distinguish historical direction from current evidence while keeping the visible layout compact. All spread, order-book depth and minimum-volume conditions have been removed from this RS screen. Direction compares price charts and momentum only; it is not an assessment of execution conditions.

## Validation status and acceptance plan

Automated checks cover causal closed-bar cutoffs, the hard 12h cutoff, recent-heavy weighting, high-beta counterexamples, missing/gapped data, invariance to old-history mutations, data staleness, finite Monte Carlo p-values, universe adjustment, request deduplication, queue priority and shared cooldown. These checks establish mechanics, not predictive accuracy.

**Not completed:** a profitability backtest, an out-of-sample accuracy estimate, or calibration of the Balanced cutoffs. The live intraday cache cannot establish performance across several market regimes. Do not call live candidate counts an accuracy test.

Before claiming predictive quality: freeze this specification; collect a point-in-time universe including delisted contracts; run chronological walk-forward evaluation across bull, bear and sideways periods; separate model-development and untouched evaluation periods; compare against no-filter and simple relative-return baselines; measure 1h/4h/12h forward BTC-relative returns, absolute returns, false-direction rate, coverage and turnover separately; use non-overlapping evaluation blocks or dependence-aware intervals; include fees, spread, slippage and funding for any trading-return claim. Report both rejected and accepted cases and confidence intervals, with no parameter choice based on the final holdout. A small or empty qualified set is allowed, but must be distinguished from loading, stale data or service failure.