← archive

research

Catalyst Detection Engine v2: Peer Review & Architecture Spec

May 2026

↩ Part of the Catalyst Radar project — see the shipped engine and the v1 optimization brief that preceded this review.

The v1 stack is structurally sound but leaves 30–60% of available IR on the table and contains three load-bearing assumptions that the empirical record contradicts. The single highest-ROI architectural change is replacing equal-weight AND/OR rules with a LightGBM meta-label gate trained on triple-barrier labels, which lifts precision 5–10pp in published reproductions on both ES futures (Hudson & Thames 0.48→0.54) and BTC (Quang Khải reproduction 0.63 precision, 0.73 recall). The single highest-ROI signal addition is CryptoBERT news classification for Pipeline B (~6pp accuracy lift over keyword RSS). Nitter is dead, X scraping is broken, DefiLlama unlocks API is paywalled, Coinglass liquidation heatmaps are paywalled, CryptoQuant has no free API, and DefiLlama has no per-token CEX flow endpoint — five free-tier assumptions in v1 that no longer hold in 2026. Realistic precision/recall floors on free data: 55–65% precision and 30–45% recall at the v1 horizons; the v1 target of ≥65% precision is achievable only on T1 majors with macro-event triggers and on T2 listing/ETF events, not as a universe-wide floor.

The peer review below is organized as the eight required deliverables, with empirical findings overriding theoretical expectations per the brief’s empirical-override clause.


1. Free data source manifest

Reliability ratings: A = production-grade native; B = stable but degraded (rate limits, scraping required); C = works but unofficial/fragile; F = dead, paywalled, or unusable on free tier.

Orderflow (Pipeline A)

SignalBest free sourceEndpointUpdateRate limitAuthReliabilityLighter coverage
Volume z-scoreLighter native WSwss://mainnet.zklighter.elliot.ai/stream trade:{market_id}50ms batchNot publishedNone publicANative primary
OI deltaLighter native + Coinalyze for CEXLighter market metadata; api.coinalyze.net/v1/ (40 req/min, free key)5 min40/minFree keyA native, B aggregatedNative + free CEX aggregator
Funding rateLighter native + Binance/Bybit/OKX/HL public/api/v1/fundings (Lighter); /fapi/v1/premiumIndex (Binance); HL /info metaAndAssetCtxsPer-block / 1minPublic weightsNoneANative
Cross-exchange basisLighter mark + HL /info + Binance premiumIndexWS streamsReal-timePer-venueNoneANative — Lighter mark price queryable
BB width / ATR / RVDerived from Lighter candlesGET /api/v1/candlesticksPer-barNoneNoneADerived
CVDLighter trade tape WS (taker side flagged)trade:{market_id}TickNoneNoneANative
Liquidation eventsHL liquidations WS, Binance forceOrder WSWSReal-timePer-venueNoneBRealized only — synthetic heatmap requires building
Liquidation heatmap (forward-looking)Coinglass paid onlyn/an/an/an/aFBuild synthetic from realized liq events + leverage prior (~12–20hr); accept lower recall

Catalyst (Pipeline B)

SignalBest free sourceMethodReliability2026 status
News RSS aggregationCoinDesk + CoinTelegraph + Decrypt + The Block + Crypto Briefing + BeInCrypto + Bitcoin Magazinefeedparser fan-in every 1–5 minAAll live; combine with CryptoBERT classification
Token unlocksDefiLlama dashboard scrape + open-source DefiLlama/emissions-adapters repo forkHTML scrape + local cronA−DefiLlama /api/emissions is Pro-only ($300/mo) — v1 assumption wrong; workaround is free fork
Binance listingsbapi/composite/v1/public/cms/article/catalog/list/query?catalogId=48JSON pollBCloudflare-protected, 403s under load; backup: HTML scrape binance.com/en/support/announcement/c-48
Coinbase listingsDiff-poll api.exchange.coinbase.com/products for new symbolsRESTAPer-listing blog discontinued; @CoinbaseAssets X is the official channel — product-list diff is the only clean free path, with minutes of latency
Upbit listingsapi-manager.upbit.com/api/v1/announcementsJSONA−Korean parsing required; produces documented “listing pump” + Kimchi premium
OKX listingsokx.com/api/v5/support/announcements?annType=announcements-new-listingsJSONA−Officially documented
Bybit listingsapi.bybit.com/v5/announcements/index?type=new_cryptoJSONA−600/5s
GitHub commit velocityapi.github.com/repos/{owner}/{repo}/commits + /stats/commit_activityRESTA5,000/hr authenticated
Reddit mention velocityoauth.reddit.com Data APIOAuthB100 QPM free non-commercial; Pushshift dead
Telegram public channelsTelethon (MTProto) or t.me/s/{channel} HTMLAPI clientB+Critical for CEX listing leaks; arXiv 2204.12929 documents +9.5% return at x=60h pre-pump
X/Twitter mention velocityNo clean free pathNitter dead, X v2 free read killed, twscrape requires burner accounts + proxies (TOS violation, account churn)FLargest free-tier gap in 2026. Substitute basket: Bluesky Jetstream + Farcaster Neynar + Reddit + Telegram
Bluesky firehosewss://jetstream2.us-east.bsky.network/subscribeWSA for tech / C for crypto signal densityFree, real-time, unlimited; small crypto-trader population
Farcaster mentionsNeynar api.neynar.com/v2/farcaster/cast/searchRESTBCrypto-native users; free tier viable for ~100 tickers hourly
Macro calendar (FOMC)federalreserve.gov/monetarypolicy/fomccalendars.htmHTML scrapeAAuthoritative
CPI / NFPBLS RSS + Treasury directRSSAFree
ETF decisionsSEC EDGAR cgi-bin/browse-edgar?action=getcurrent&type=19b-4RSSAFree, official
CoinMarketCal eventsdevelopers.coinmarketcal.comOAuth RESTBFree tier exists; quality variable

On-chain / flow (Pipeline C)

SignalBest free sourceReliabilityNotes
Stablecoin (USDT) net inflow to CEXEtherscan v2 + Tronscan + custom CEX wallet listAOnly on-chain signal with peer-reviewed 1–2h predictive power on BTC/ETH (Chi/Chu/Hao 2024)
Per-token CEX exchange netflowEtherscan v2 + Routescan + Solscan + custom CEX listBDefiLlama does NOT expose per-token CEX flows on free tier — v1 assumption wrong
ETH net inflow → ETHSameANegative predictor 1–6h (Chi et al. 2024)
BTC net inflow → BTCSameF as predictorNo significant intraday predictive power (Chi et al. null result) — drop
Bridge volume per chainDefiLlama bridges.llama.fiAMulti-day narrative; not intraday
DEX vs CEX volumeDefiLlama + CoinGecko Demo (30/min, 10k/mo)ADays timescale (Makarov-Schoar) — regime filter only
Whale wallet trackingCielo free (250 wallets, no API analytics) + Whale Alert Twitter + custom Etherscan v2 watchlistsBTrue smart-money labels remain paid (Nansen $150/mo)
Hyperliquid public stateapi.hyperliquid.xyz/info + WSABest-in-class free perp-DEX feed; primary cross-DEX whale tracker
Lighter L1 USDC depositsEtherscan v2 on 0x3B4D794a66304F130a4Db8F2551B0070dfCf5ca7AUnder-researched niche; pre-listing leak detector
Token holder concentrationEtherscan v2 (top-holder API moved to Pro) + custom tokenTx aggregationCMulti-day signal — drop for intraday
CryptoQuant netflowWebsite charts only — no free APIFv1 assumption wrong; use as daily regime context only

Five v1 free-source assumptions that no longer hold in 2026: Nitter (dead), DefiLlama unlocks API (paywalled), DefiLlama per-token CEX flow (does not exist), Coinglass liquidation heatmap API (paywalled), CryptoQuant API (paid only).


2. Signal performance table — standalone baselines

Honest framing: precision = P(≥2% move within horizon | signal fires), recall = P(signal fires | move occurred). Most cited literature uses different metrics (IC, Sharpe, R², F1); the numbers below translate or extrapolate cautiously and are flagged ⚠️ when extrapolated. Base rate of ≥2% move in arbitrary 4-hour window for a typical perp ≈ 8–12%.

Top single signals ranked by composite score (precision × recall × 1/latency × free-tier reliability)

RankSignalBest horizonPrecisionRecallLatencyEffortSourceConfidence
1Cross-exchange basis spread (Lighter–HL or Lighter–Binance >1.5σ)1–2h0.45–0.600.25–0.40<5s8–12hAlexander et al. 2024; He et al. arXiv:2212.06888High
2Tier-1 CEX listing announcement (Binance/Coinbase/Upbit)1–2h0.65–0.75 ⚠️event-boundseconds16–24hRen & Heinrich; Messari Coinbase Effect; CryptoNinjas/StoribleMedium-High; T2/T3 only
3USDT net inflow to CEX → BTC/ETH1–2h~0.52–0.55low (top-decile only)~10 min20–30hChi/Chu/Hao 2024 arXiv:2411.06327Medium-High
4OFI (Cont/Kukanov/Stoikov) on Lighter<30 min0.40–0.550.20–0.35<1 min6–10hCont et al. arXiv:1011.6402; Springer Digital Finance crypto replicationHigh
5Funding rate extreme (>95th or <5th pct, 30d) into binary catalyst6–24h0.45–0.550.20–0.3560s3–5hPresto Research; He et al. 2022; empirically 7/8 FOMC pre-positioning hitMedium
6Volume z-score >3 (5min)1–2h0.30–0.45 majors / 0.20–0.30 memes0.50–0.65<1 min6–10harXiv 2412.18848 pump detection 55.81% top-5 hit-rateMedium; ⚠️ wash-trade-corrupted on T3
7Token unlock (>5% supply, team category)6–24h0.55–0.65 DOWN0.40–0.55scheduled8–16hKeyrock 16k unlocksMedium; ⚠️ much weaker than v1 assumes for monthly cliffs
8**OI delta + flat price (>2σ,Δprice<0.3%)**6–24h0.40–0.550.25–0.401–5 min4–8h
9VPIN >0.7 (volume-bucketed toxicity)1–2hregime indicatorn/a<1 min8–12hKitvanitphasu et al. 2025 ScienceDirect; Easley/López de Prado/O’HaraMedium; use as risk filter, not alpha
10CVD divergence1h0.30–0.450.15–0.25<1 min6–10hCont et al.; Sirignano-Cont 2019High at <1h, decays
11CryptoBERT news sentiment + entity-link2–6h0.55–0.62 ⚠️low for ≥2%1–5 min24–40hACSA 2023 (CryptoBERT XL F1 0.59); MDPI 2024Medium
12Telegram pump-channel monitoring1–2h0.50–0.60 ⚠️lowseconds16–32harXiv 2204.12929 (+9.5% at x=60h)Medium for risk avoidance
13BB width compression (<5th pct)6–24h0.40–0.55 (direction-agnostic)0.55–0.70<1 min2–4hQuantified StrategiesMedium; needs direction confirmation
14Reddit mention velocity2–6h0.53–0.58 ⚠️ BTC/ETH onlylow for long tailminutes16–24harXiv 1907.00558Medium for majors, low for tail
15GitHub commit velocity6–24h0.52–0.55 ⚠️ direct; better as filterlowhourly8–16hPhillips & Gorse 2018Low-Medium
16CryptoQuant BTC exchange netflow1–7 dayn/a intradayn/a24h+Glassnode Insights 2025Horizon mismatch — drop for sub-24h
17Holder concentration / new whale detectionweeksn/a intradayn/ahours40h+IntoTheBlockHorizon mismatch — drop
18Bridge inflow per chain1–7 dayregimen/ahourly4hDefiLlama dashboardsHorizon mismatch — narrative filter only
19Stablecoin aggregate supply growthdays-weeksmacron/adaily3hCoinMetrics SOTNHorizon mismatch — daily regime only
20Liquidation cluster proximity (synthetic free)1–2h0.30–0.500.10–0.20 (low recall on free proxy)<1s12–20hGate market signals; coinglass methodology paywalledLow — structurally weaker than paid

Signals to drop from v1: BTC exchange netflow (Chi et al. null result intraday), holder concentration (multi-week), bridge volume per chain (multi-day), CryptoMarketCal events as standalone (community-voted noise), generic “cross-exchange spread” without DEX-CEX framing.

Signals to add that v1 is missing: OFI (Cont/Kukanov/Stoikov), VPIN, CryptoBERT sentiment, Telegram pump-channel scrape, Hyperliquid whale-fill cross-reference, ETF flow direction (substitute for stablecoin inflow at T1).


3. Top three stacks per pipeline

Stacks are ranked by composite score = precision × recall × (1/latency_min) × free_reliability. Each stack has a plain-English explainer plus exact thresholds.

Pipeline A — Orderflow (1–2h, scan every 2–3 min)

Stack A1 (highest conviction): Basis dislocation + OFI confirm + funding regime. What it does in plain English: fires when Lighter’s perp price diverges from Hyperliquid/Binance by more than 1.5 standard deviations of recent spread, the Lighter L2 order-flow imbalance confirms direction, and funding sits at a non-extreme percentile so the move isn’t already crowded. This catches venue-leading informed flow.

  • Lighter–HL basis z > 1.5σ (rolling 24h)
  • OFI sign matches basis direction, OFI 5-min EMA |z| > 1.5
  • Funding rate between 30th and 70th percentile (avoids crowded carry trades)
  • Estimated precision 0.55–0.65, recall 0.20–0.30

Stack A2: Volume + OI delta + funding extreme reversal. What it does: identifies leverage build-up that’s about to flush — high volume z, OI ramping with flat price, extreme funding percentile signaling crowded positioning. This is the classic “squeeze setup” but with a CVD confirmation requirement to avoid v1’s empirically inverted assumption (see counterintuitive finding #2).

  • Volume z > 3 (5-min vs 30d)
  • OI delta > 5% in 1h with |price Δ| < 0.3%
  • Funding > 95th or < 5th percentile (30d)
  • Required filter: spot OBV z > 0.5 in matching direction (kills basis-trade FP)
  • Estimated precision 0.45–0.55, recall 0.20–0.30

Stack A3: VPIN + liquidation proximity + BB squeeze release. What it does: detects toxic order flow regime preceded by volatility compression and a liquidation cluster nearby — the classic “vol-crush before vol-expansion” pattern, gated by VPIN >0.7 to filter out pure mechanical compression.

  • VPIN > 0.7 (volume-bucketed)
  • BB width < 10th percentile (20-period, 30d distribution)
  • Within 1% of synthetic liquidation cluster
  • Estimated precision 0.40–0.50, recall 0.30–0.45

Pipeline B — Catalyst (2–6h, scan every 60 min + event-driven push)

Stack B1 (highest conviction): T1 CEX listing announcement + spot CVD confirmation. What it does: listing announcements on Binance/Coinbase/Upbit produce documented +29% to +73% pumps, but pre-pump basis arb FPs are common. Requiring spot CVD confirmation within ±6h kills the listing-arb false positive (FP-7) that makes OI ramp without directional intent.

  • Binance/Coinbase/Upbit listing detected via API or RSS within 30 min
  • Spot CVD z > 1.0 in same direction within 30 min of announcement
  • Estimated precision 0.65–0.75, recall ~event-bound

Stack B2: CryptoBERT-classified bullish news + funding regime + ticker velocity. What it does: pre-trained CryptoBERT XL classifies news as bullish/bearish (58.5% accuracy on crypto vs FinBERT 52.7%); combined with non-crowded funding and a Reddit/Telegram mention spike for the named ticker. The mention spike confirms the news is being absorbed by the market in real-time.

  • CryptoBERT score > 0.7 bullish or < 0.3 bearish
  • Reddit + Telegram mention z > 2.5 (24h baseline) for ticker
  • Funding 30th–70th percentile
  • Hours-to-FOMC > 4 (kill macro-event interference)
  • Estimated precision 0.55–0.65, recall 0.20–0.35

Stack B3: Token unlock + recipient = team/investor + pre-drift. What it does: targets the subset of unlocks that empirically dump (>5% supply, team/VC recipient, with pre-event drift not already negative). Per Keyrock’s 16k-event study, ecosystem unlocks average +1.18% — they must be excluded.

  • Unlock event T-2 days
  • Unlock size > 5% circulating supply
  • Recipient = team or VC (not ecosystem/community)
  • 14-day pre-drift not already < −5% (already priced in)
  • Direction = SHORT
  • Estimated precision 0.55–0.65, recall ~event-bound

Pipeline C — Whale + Narrative (6–24h, keep 4h cadence)

Stack C1 (highest conviction): Hyperliquid whale fills + Lighter mirror + funding regime. What it does: the top-PnL wallets on Hyperliquid (free public API) leading large positions in tokens also listed on Lighter is the cheapest cross-perp-DEX whale tracker available in 2026. Mirroring direction-agnostic flows misses; require the same direction to hit Lighter within 1–4h.

  • Top-100 HL trader takes new position > $1M in matching ticker
  • Lighter sees same-direction OI delta within 1–4h
  • Funding not at extreme that would imply crowded
  • Estimated precision 0.55–0.65, recall ~position-event-bound

Stack C2: USDT exchange inflow + ETH/BTC pair + ruptures change-point. What it does: aggregates USDT net inflow to top CEXs (the only on-chain signal with peer-reviewed 1–2h predictive power on BTC/ETH per Chi et al. 2024), confirmed by a structural break in OI series detected by ruptures PELT. This is regime-shift detection rather than threshold detection.

  • Aggregated 1h USDT inflow to top-5 CEXs > 95th percentile (30d)
  • ruptures change-point detected on log-OI series in last 4h
  • For ETH: signal = inflow positive → SHORT (per Chi et al. negative predictor for ETH 1–6h)
  • For BTC: signal = inflow positive → LONG
  • Estimated precision 0.50–0.60, recall 0.15–0.30

Stack C3: Sector rotation + ETF flow + isolation forest cross-section. What it does: identifies the sector currently winning capital rotation (AI / RWA / DePIN / L2) via ETF flow and price-momentum, then runs an isolation forest on the cross-section of ~100 perps to flag tokens with anomalous volume/OI/funding combos within that hot sector. Replaces v1’s “sector pulse” hand-coded indices with an actually-implementable cross-sectional anomaly score.

  • Sector ETF flow > 75th percentile (T1 only)
  • Isolation forest contamination=0.05 anomaly score on (vol z, OI z, funding pct, basis z) features
  • Token belongs to leading sector
  • Estimated precision 0.45–0.55, recall 0.30–0.45

4. Per-tier optimization matrix

The optimal stack differs materially by tier because market structure differs materially. Weights below are recommended priors for the meta-label LightGBM training set; final weights should be learned empirically on rolling 60-day windows.

TierPipeline A weightPipeline B weightPipeline C weightJustification
T1 Majors (BTC/ETH/SOL/BNB/XRP)0.450.350.20Macro-event-driven 2–6h moves dominate per empirical reverse-engineering (FOMC ×3, CPI ×2, geopolitical ×6 of 24 documented moves). Funding extremes pre-event reliably forecast post-event direction. Whale flow is slower; stablecoin inflow has peer-reviewed 1–2h predictive power for BTC/ETH but recall is low. ETF net flow last 24h substitutes for “stablecoin inflow” — empirically much more observable in 2026 macro regime.
T2 Mid-caps (LDO/ARB/OP/INJ/HYPE/etc)0.250.500.25Listing announcements + ETF filings + protocol revenue narratives drive 5–20% moves (HYPE +108% from low; AIGENSYN +250% on Binance Alpha + Coinbase listing). Catalyst weight is highest here. Whale concentration matters for pre-listing leak detection (Telegram). Orderflow is cleaner than T3 but noisier than T1.
T3 Memes (WIF/PEPE/BONK/FARTCOIN/MOG)0.20 (wash-corrupted)0.300.50Whale wallet concentration is both the entry signal and the exit warning (FARTCOIN $145M TWAP build → +250% pump → −50% liquidation; RaveDAO +3,765% → −95% on ZachXBT exposure). Volume z-scores are corrupted by wash trading (Cong et al. ≥70% of unregulated CEX volume is wash). Hard gate: T3 alerts require Benford χ² < 15.5, round-clustering < 0.35, and >100 unique trader addresses in 24h. Social-source-quality filter (require credible KOL/onchain-detective post, not just mention count).

Cross-tier filters that override tier weights:

  • Macro event ±30 min (FOMC/CPI/NFP) → suppress all non-T1 alerts
  • Cross-venue heartbeat lag >30s OR price dispersion >2% → global 60-min freeze (avoids Oct-2025-style outage FPs)
  • Within 24h of monthly/quarterly Deribit expiry AND |spot − max-pain|/max-pain < 2% → suppress momentum alerts on BTC/ETH

5. Cross-pipeline conviction multipliers

Empirical anchor: combining alpha streams in published research lifts IR by 1.38× to 1.7× (RavenPack: 4.35→6.0; multiple multi-factor crypto studies). Theoretical Bayesian upper bound assuming conditional independence with each pipeline at precision 0.20–0.30 and FP rate 0.07: 3-pipeline confluence reaches ~0.75 posterior, but real-world correlation (ρ ≈ 0.2–0.4) deflates to ~0.55–0.65.

Active pipelinesConviction multiplier vs single-pipeline baselineSource / Rationale
A only (orderflow)1.0×OFI predictive but decays in <30 min (Cont et al.)
B only (catalyst)0.9×News absorbs in ~45 min; high FP without confirmation (SJFA 2025)
C only (whale/onchain)1.1×Slower; higher base precision when fired
A + B1.7×Independent timescales; orderflow confirms catalyst absorption
A + C1.8×Smart money + tape; historically strongest combo (Apr 2026 BTC rally pattern)
B + C1.5×Both slow; better for swing setups than entry timing
A + B + C2.5–3.0×Rare; per Grinold √3 with ρ≈0.3 implies ~1.6× IC, ~2.5× edge after Bayes

Operational rule: treat these as priors for the meta-label LightGBM confluence feature, not hard score multipliers. The model should learn the actual uplift from labeled history; the priors above seed the cold-start period.


6. False-positive filter rules

Twelve documented patterns with detection logic. The first six are hard gates (suppress regardless of composite score); the rest are soft penalties applied to the score before meta-label evaluation.

#PatternDetection conditionActionTier sensitivity
1Basis-trade OI ramp (Ethena/cash-and-carry style)OI z > 2 ∧ |spot OBV z| < 0.5 ∧ funding > 0.0001Hard suppress long-momentumT1 ≫ T2
2Funding flip mechanicalsign-flip ∧ |spot ret 1h| < 0.3% ∧ vol z < 1Hard suppress reversalT1, T2
3MM block / inventory rebalancevol z > 2.5 ∧ trade-count z < 1 ∧ markout-5min ≈ 0Hard relabel as block; kill momentumT2 ≫ T3 ≈ T1
4Cross-venue outagevenue HB lag > 30s OR cross-venue px disp > 2%Hard global 60-min freezeAll
5CEX listing arb spilloverwithin ±6h Tier-1 listing ∧ OI z > 2 ∧ basis ≈ 0 ∧ no spot CVD confirmHard require spot CVD confirmationT2 ≫ T3
6Wash trading T3Benford χ² > 15.5 OR round-cluster > 0.35 OR Pareto α ∉ [1,2] OR <100 unique traders/24hHard suppress entirelyT3 only
7Pre-priced news / sell-the-news14-day pre-drift > 10% ∧ within ±2h known calendar eventSoft −40%All, especially T1
8Stop-hunt / liquidity sweepwick/range > 0.6 ∧ candle closes back inside prior rangeHard invalidate breakout, allow fade T+1All
9Liquidation hunt one-sidedliq z > 3 ∧ one-side > 85%Soft −50% for 15 min in same directionT1 ≫ T2
10Token unlock priced-inunlock < 1% circ OR recipient = ecosystem OR pre-drift < −5%Hard suppress short alertsT2 ≫ T3
11Stablecoin mint not deployedexchange-inflow / mint < 50% in 24hSoft −30% on risk-on signalsT1
12TWAP / max-pain pintrade-size CV < 0.3 ∧ interval CV < 0.3 OR |spot − max-pain|/max-pain < 2% within 24h of expirySoft −40% / hard pre-expiry suppressT1 (BTC/ETH)

These twelve rules go in the Python prefilter stage before meta-label evaluation. The meta-label model then trains on the residual events (post-filter), so the LightGBM doesn’t waste capacity learning patterns the deterministic filters already handle.


7. Final v2 architectural spec

What survives from v1

  • Three-pipeline parallel architecture — empirically validated by horizon-specific signal half-lives (OFI <30min, news ~45min, whale flow 6–24h+)
  • 4h Pipeline C cadence — correct given signal timescales
  • Three-tier universe segmentation — empirically validated by distinct pre-move signatures per tier
  • Score-band gating (>70 high-conviction, 50–69 watchlist, <50 suppress) — kept as transparent fallback when meta-label model is being retrained or its OOS AUC drops below 0.55

What gets cut

  • Equal-weight 0.33 layer combination → replaced by LightGBM meta-label gate
  • 5-minute Pipeline A cadence → tightened to 2–3 min where infra allows; OFI dies at 15–30 min
  • DefiLlama unlock calendar assumption → fork open-source emissions-adapters or scrape dashboard
  • Nitter-based X/Twitter scraping → dead; replaced by Bluesky + Farcaster + Reddit + Telegram basket
  • Coinglass free-tier liquidation heatmaps → build synthetic from realized liq events
  • DefiLlama per-token CEX flows → does not exist on free; build via Etherscan v2 + custom CEX wallet list
  • CryptoQuant netflow API → no free tier; use website chart scraping for daily regime context only
  • Holder concentration as intraday signal → drop (multi-week timescale)
  • Bridge volume per chain as intraday signal → drop (multi-day timescale)
  • BTC exchange netflow as intraday predictor → drop (Chi et al. null result)
  • CoinMarketCal events as standalone → demote to lead-generator only

What gets added

  • LightGBM meta-label gate trained on triple-barrier labels (TP=2%, SL=1%, vertical=horizon), retrained walk-forward weekly with isotonic calibration. Single highest-ROI architectural addition; +5–10pp precision in published reproductions.
  • Cont/Kukanov/Stoikov OFI computed on Lighter L2 stream — interpretable, ~30 LOC, peer-reviewed
  • VPIN (volume-bucketed toxicity) — risk filter for liquidation-cascade regimes; ~50 LOC
  • CryptoBERT (kk08/CryptoBERT or ElKulako/cryptobert) for news classification — F1 0.59 on crypto vs FinBERT 0.53; runs free on CPU
  • Hyperliquid whale-fill cross-reference — best free cross-perp-DEX whale tracker
  • Lighter L1 USDC deposit anomaly detector — under-researched niche signal; cheap to build
  • Telegram pump-channel monitoring — best free leading indicator for CEX-listing leaks per arXiv 2204.12929
  • Isolation Forest as cross-sectional pre-filter — sklearn one-liner; flags interesting tokens
  • ruptures change-point detection on Pipeline C OI series — regime-shift feature, not entry trigger
  • 3-bucket realized-vol regime context (instead of HMM, which is overhyped at sub-daily) — captures 80% of regime benefit at 1% of the cost
  • Triple-barrier labeling via mlfinlab.labeling — eliminates fixed-horizon heteroscedasticity bug

What gets explicitly skipped

  • HMM regime detection (Hidden Markov Models) — overhyped at sub-daily horizons; simple vol-quantile bucket suffices
  • LSTM / Transformer sequence models — do not beat LightGBM on tabular short-horizon classification (arXiv 2309.11400)
  • Hawkes processes — real edge but heavy implementation; marginal lift over OFI for binary classification; defer to v3
  • River online learning — weekly LightGBM retrain captures 90% of drift-adaptation benefit with much better stability
  • Bayesian model averaging — operationally noisy with N=3 signals and short crypto histories; logistic regression with isotonic calibration delivers 95% of benefit at 5% of cost

Final cadence per pipeline

Pipelinev1 cadencev2 cadenceRationale
A (orderflow)5 min2–3 min where infra allows; 5 min with exponential recency-weight (τ=10min) otherwiseOFI dies at 15–30 min
B (catalyst)60 min60 min polling + event-driven push on listings/unlocks/macroNews absorbed in ~45 min; listing pumps complete in 24–72h
C (whale/narrative)4 hrKeep 4 hrSignal half-life 6–24h; faster polling adds noise

Realistic precision/recall floors on free data

  • Universe-wide: 55–65% precision, 30–45% recall — the v1 ≥65% precision target is not achievable as a universe floor on free data
  • T1 majors with macro-event triggers (Pipeline B): 65–75% precision attainable
  • T2 listing/ETF events (Pipeline B): 65–75% precision attainable
  • T3 memes (any pipeline): 45–55% precision realistic ceiling due to wash trading, oracle thinness, rug-pull risk

8. Build effort estimate

ComponentHoursNotes
Lighter native WS ingestion (trades, OI, funding, mark, candles)16Async websocket client + rolling state
Cross-venue mark/funding ingestion (Binance, Bybit, OKX, HL via ccxt)12ccxt covers all four; Lighter not in ccxt registry — separate client
Coinalyze free aggregator integration6OI/funding/long-short ratio cross-CEX
Pipeline A signal engine (vol z, OI z, funding pct, basis, BBW, ATR, CVD, OFI, VPIN)60Most are <30 LOC each; OFI and VPIN are the largest
Synthetic liquidation heatmap (realized liq events + leverage prior)16Aggregate HL/Binance/Bybit/OKX liq WS
News RSS fan-in + entity linking (spaCy + ticker gazetteer)16feedparser + custom disambiguation
CryptoBERT inference pipeline8HuggingFace transformers on CPU
Token unlock dashboard scrape + emissions-adapters fork12Local cron of forked DefiLlama repo
Listing announcement pollers (Binance, Coinbase product diff, Upbit, OKX, Bybit)16Multi-source with retry/backoff
GitHub commit velocity per protocol8Mapping repos via DefiLlama metadata
Reddit OAuth + mention-velocity z-scores12100 QPM budget across ~100 tickers
Telegram channel scraping (Telethon)16FloodWait handling + channel curation
Bluesky Jetstream + Farcaster Neynar mention pipelines16WS + REST
Macro calendar (FOMC/CPI/NFP/ETF EDGAR) ingestion8RSS + scheduled polling
Stablecoin CEX inflow pipeline (Etherscan v2 + Tronscan + custom CEX wallet list)24Wallet list curation is the bulk
Hyperliquid whale-fill tracker (top-100 PnL leaderboard scrape + position diff)20HL public API + state diffing
Lighter L1 USDC deposit anomaly detector12Etherscan v2 on deposit contract
Bridge / DEX vs CEX / sector flow pipelines (DefiLlama)12Multiple endpoints, simple aggregation
Isolation Forest cross-sectional pre-filter4sklearn
ruptures change-point on OI4Library call
Triple-barrier labeling pipeline12mlfinlab.labeling integration
LightGBM meta-label model + walk-forward retrain harness24Weekly scheduled retrain + isotonic calibration
Twelve hard-gate / soft-penalty filter rules24Each is small, but needs unit tests
Wash-trading filter (Benford + rounding + Pareto + uniqueness)12Weekly recompute per T3 token
Score combiner + alert payload generator8Final orchestration layer
Backtesting harness (synced historical data + replay)32Required for meta-label training and v2 validation
Monitoring, logging, alert dispatch16Required for production
Total to live alerts~404 hours~10 weeks at 40hr/wk; ~6 weeks at 70hr/wk with LLM-assisted dev

Single VPS (~$5–10/mo) acceptable as compute infrastructure under “free data” constraint. Total external API spend: $0/mo.


9. Three most counterintuitive findings

1. OI ramp with flat price is bearish, not bullish, in 2026’s regime. v1 treats this as a “squeeze setup” likely to resolve directionally upward. Empirical 2026 evidence inverts this: CryptoQuant’s May 2026 piece documents that BTC’s 12.7% April gain was “perpetual-futures-driven; spot demand contracted” and each prior instance preceded a fade. The Apr 17 BTC rally, the Apr 14 SOL rally, and HYPE’s spot-CVD-divergence-during-rally pattern all confirm this. Implication: v1’s OI-ramp signal must be paired with a hard spot-OBV-confirmation gate; without spot confirmation, treat OI ramps as basis-trade FPs (FP-1) not directional setups. This is the single largest empirical contradiction to v1 in the dataset.

2. Token unlocks are not reliable bearish catalysts at the cadences v1 assumes. v1 weights unlocks as a high-conviction Pipeline B signal. Keyrock’s 16,000-event study and Tokenomist data both show: pre-event drift starts 30 days early, the actual unlock day has minimal price impact, ecosystem/community unlocks average +1.18% (positive), and monthly cliff schedules (e.g., ARB) get absorbed after the first 2–3 events. Only large team/VC unlocks with non-already-negative pre-drift are reliably bearish. Implication: unlock signal needs recipient classification, size threshold, and pre-drift gate — without these, it’s a 50-50 noise signal that v1 weights as high-conviction.

3. Headline classification is the missing necessary condition v1 doesn’t have. No single technical signal in v1’s spec hit ≥80% frequency across the 24 documented Q1–Q2 2026 moves. The closest thing to a necessary condition was “News headline OR extreme funding” at 92% frequency. Geopolitical headlines alone (Iran ceasefire, Strait of Hormuz, Trump-Iran address) produced ≥3% moves with 100% directional accuracy in the dataset. v1’s “RSS aggregation” treats headlines as triggers but not as classified directional signals — a CryptoBERT classifier on headline corpus is the missing necessary condition, not a nice-to-have. Combined with the Nitter-is-dead reality, this means Pipeline B in v2 is fundamentally a news-classification pipeline with social-velocity confirmation, not a social-velocity pipeline with news flavoring.

Bonus finding: Lighter is a follower, not a leader. Lighter’s $1B 24h volume vs Hyperliquid’s $180B/30d means Lighter prices follow Hyperliquid flow with measurable lag. v1 treats Lighter as the primary venue for signal generation. The empirical record suggests v2 should use Hyperliquid orderflow as the leading indicator and Lighter as a confirmer/divergence check. This inverts the v1 mental model and changes which API is primary in Pipeline A.


Conclusion

The v2 spec keeps v1’s three-pipeline / three-tier topology but rebuilds the combiner around a LightGBM meta-label gate, replaces five paywalled or dead free-data assumptions with working alternatives, adds the four signals v1 was missing (OFI, VPIN, CryptoBERT, Hyperliquid whale tracking), and inverts two load-bearing v1 assumptions (OI-ramp directionality, unlock bearishness) that contradict the 2026 empirical record. Realistic precision floors on free infrastructure are 55–65% universe-wide, with 65–75% achievable on T1 macro events and T2 listing events specifically. Total build effort to live alerts: ~400 hours. Total external API spend: $0/month.

The biggest remaining gap is X/Twitter — Nitter is dead, X v2 free read is killed, and the Bluesky+Farcaster+Reddit+Telegram substitute basket is structurally weaker than what was possible in 2023. If this gap proves binding in production, the cheapest paid escape hatch is Apify Twitter actors at $30–50/mo metered, which is a strict violation of the zero-paid-API constraint but worth flagging as the single highest-leverage paid upgrade if the basket underperforms.