research
Catalyst Radar v1: Architecture Optimization Brief
May 2026
↩ Part of the Catalyst Radar project — see the shipped engine and the rest of its evolution.
The 92-hour build target underprices the right architecture by ~40%. The v1 spec has three structural flaws — 30-min uniform cadence, single-Haiku-per-ticker, and 24h%-led ranking — that must be fixed before signal-to-noise reaches the ≥75% catalyst-attribution and ≥80% mover-recall targets. The fixes are: a hybrid event-driven + tiered-periodic loop, a two-stage embedding-prefilter pipeline that reserves Haiku for natural-language rationale only, and a vol-normalized max-window-pop score with funding-rate and BTC-beta residualization. Build revised to ~125–140 hours; recurring spend stays under $10/mo. The leveraged-only universe filter is the single best decision in the v1 spec because it makes funding rate, OI velocity, and liquidation cascades universal signals across crypto, equity, commodity, and FX perps — exploit that.
1. Free catalyst source manifest by asset class
Reliability scale: A = official, low-latency, machine-readable, stable. B = official but HTML/scrape or moderate latency. C = third-party scrape, ToS-grey or fragile. F = paywalled or dead.
Crypto perps — top 5 by ROI
| Source | Endpoint | Freq | Limit | Reliability |
|---|---|---|---|---|
| Bybit announcements (official) | api.bybit.com/v5/announcements/index?type=new_crypto | Real-time | 600/5s | A |
| Upbit notices (Korean) | api-manager.upbit.com/api/v1/notices/search | Real-time | ~6/min safe | B |
| Binance bapi listings | binance.com/bapi/composite/v1/public/cms/article/list/query?catalogId=48 | Real-time | Undocumented; 1/30s + UA rotation | B (officially “not for public use”) |
| Coinalyze (funding/OI/liq) | api.coinalyze.net/v1/... | 1-min | 40/min FREE | A |
| DefiLlama emissions | api.llama.fi/emissions | Daily | None practical | A |
Supporting tier: CoinDesk RSS (coindesk.com/arc/outboundfeeds/rss/), The Block RSS, Decrypt RSS, Etherscan v2 (5/sec, 100k/day free), CoinGecko Demo categories (30/min free key), Whale Alert public Telegram (live API is paid; archive free 2026-Q1).
Critical paywall gaps with no clean free workaround: Coinglass ($29/mo minimum), Twitter/X firehose (no usable free API), Nansen/Arkham smart-money labels (must self-curate from public Dune dashboards by @hildobby, @adam_tehc), live Whale Alert API. Build curated address tables by hand; this is unavoidable.
Equity perps — top 5
| Source | Endpoint | Freq | Limit | Reliability |
|---|---|---|---|---|
| SEC EDGAR Atom (per-form) | sec.gov/cgi-bin/browse-edgar?action=getcurrent&type=8-K&output=atom | Real-time | 10/sec across sec.gov; UA header mandatory | A |
| Finnhub free | finnhub.io/api/v1/calendar/earnings, /company-news | Hourly | 60/min | A |
| Yahoo via yfinance | Ticker(sym).calendar, .upgrades_downgrades, .option_chain | Daily | Aggressive throttle ~few hundred/day/IP | B |
| Farside Investors | farside.co.uk/btc-etfs/ HTML | Daily | None | A for BTC/ETH ETF flows |
| pykrx + Naver scrape (Hyundai) / yfinance ASML.AS (Euronext) | various | Daily/intraday delayed | n/a | C (ToS-grey) |
No free analyst-rating-change real-time feed exists in 2026. Use yfinance upgrades_downgrades polled hourly + 8-K/424B EDGAR scan as proxy. No free options unusual activity feed exists. Synthesize from yfinance option chain: vol/OI > 3 AND vol > 5× 20-day avg per contract. AlphaVantage free dropped to 25 req/day — unusable.
Commodity perps
| Source | Endpoint | Freq | Reliability |
|---|---|---|---|
| EIA v2 | api.eia.gov/v2/petroleum/stoc/wstk/data/ (series WCESTUS1 crude, WGTSTUS1 gasoline) | Wed 10:30 ET weekly | A |
| OPEC press RSS | opec.org/opec_web/en/feeds.htm | Event | A |
| GDELT 2.0 DOC | api.gdeltproject.org/api/v2/doc/doc?query=... | 15-min | A — universal geopolitical layer |
| Mining.com per-commodity RSS | mining.com/commodity/gold/feed/ | Hourly | B |
| Baker Hughes XLSX scrape | rigcount.bakerhughes.com/na-rig-count | Fri 1pm ET | B |
Reuters and AP have no first-party RSS in 2026. Anything labeled “Reuters RSS” is third-party scrape (rss.app/feedspot). GDELT replaces both as the universal wire layer. ACLED is weekly, ISW is editorial — neither belongs in a 30-min loop.
FX + macro
| Source | Endpoint | Reliability |
|---|---|---|
| Forex Factory XML | nfs.faireconomy.media/ff_calendar_thisweek.xml (URL has moved twice in 24mo — verify on deploy) | B |
| FRED API | api.stlouisfed.org/fred/series/observations (key series: DTWEXBGS, DGS10, DEXJPUS, T10Y2Y) | A, 120/min |
| BLS Public Data v2 | api.bls.gov/publicAPI/v2/timeseries/data/ (CPI=CUUR0000SA0, NFP=CES0000000001) | A, 500/day with key |
| Fed/ECB/BoJ/BoE/SNB/RBA RSS bundle | 6 official feeds | A |
| BIS central-bank speeches | bis.org/doclist/cbspeeches.rss | A — single firehose for global CB tone |
| Treasury Fiscal Data | api.fiscaldata.treasury.gov/services/api/fiscal_service/v1/accounting/od/upcoming_auctions | A |
| Polymarket Gamma | gamma-api.polymarket.com/markets | A — best free implied-probability source for political/regulatory catalysts |
Trading Economics free tier (100 lifetime requests) is unusable for production. Skip.
2. LLM prompt template and pipeline architecture
The single biggest finding: replace “Haiku call per ticker” with a two-stage pipeline. Embeddings + NER prefilter cuts Haiku tokens 40–60%, FinBERT direction-validator catches hallucinations Haiku cannot self-detect (Haiku 4.5 has no token-level logprobs — ECLIPSE benchmark AUC 0.59 vs 0.89 GPT-3.5). At 80 movers/day × 30 calls × 2K tokens × $1/M, this stays comfortably under $10/mo.
Required architecture:
mover detected (vol-normalized pop > threshold)
→ fetch news (RSS pool, 80–200 items)
→ spaCy NER + ticker regex (drops ~60% off-topic)
→ BGE-base-en-v1.5 embeddings, rank vs (ticker + asset_class + direction); top-5
→ BGE intra-cluster dedup (cosine ≥ 0.85 = duplicate)
→ DeepL free tier translate for KO/ZH (500K chars/mo)
→ FinBERT sentiment annotation (sidecar metadata)
→ Haiku call with Anthropic Structured Outputs (constrained decoding)
→ validator: evidence_quotes ⊂ news block? FinBERT direction ≈ Haiku direction?
→ suppression chain (§5)Use Anthropic Structured Outputs (GA Nov 2025), not prompt-only JSON. Reported approximately zero schema violations on Haiku 4.5; ~2–3% cost overhead more than offset by elimination of retry logic. Temperature 0.0, top_p=1.0. Even at 0, Claude is not fully deterministic per Anthropic SDK — accept that.
Few-shot vs zero-shot: 3 hand-curated examples (one positive, one NO_CATALYST, one stale/priced-in), diverse across asset classes. More than 3 shots regresses Haiku on financial classification (arXiv 2312.08725, 2411.02476). Universal prompt with explicit asset_class field beats per-class variants — captures most variant value at 1× maintenance cost, and Anthropic prompt-cache pricing rewards static prefixes.
Recommended prompt skeleton (copy-paste):
SYSTEM:
You are a quantitative catalyst-attribution analyst. You receive a ticker,
asset class, 24h move, trading session, and a ranked news bundle in a time
window. Identify the SINGLE most likely fundamental catalyst, OR explicitly
return NO_CATALYST. Hallucinating a catalyst is a critical failure.
Hard rules:
1. Cite verbatim quotes from <news> only. No quote → no claim.
2. Articles >24h before move_start_utc are STALE — flag is_priced_in=true.
3. trading_session ∈ {CLOSED_WEEKEND, CLOSED_INTERNATIONAL} → require news
within 6h of move_start_utc; else NO_CATALYST + thin_liquidity_flag=true.
4. Tier-3/4 sources alone → cap conviction ≤ 0.4, alert_priority ≤ P2.
<schema>{primary_catalyst, catalyst_type∈12-enum, direction, conviction,
horizon, continuation_thesis, kill_signal, is_priced_in, alert_priority,
evidence_quotes[], evidence_source_ids[], stale_news_flag, thin_liquidity_flag,
confidence_in_attribution}</schema>
USER: <asset>...</asset><news id=N published_at=... credibility_tier=1..4>...
</news><examples>3 diverse cases</examples>Anti-hallucination patterns that empirically work (rank-ordered): explicit NO_CATALYST enum (eliminates “must produce something” pressure), force-quote evidence_quotes[] with downstream substring validator, explicit credibility_tier field passed externally (LLMs cannot reliably distinguish Bloomberg from a Substack — arXiv 2502.04426), and separate confidence_in_attribution field decoupled from P&L conviction.
Multi-language handling: pre-translate, never inline. DeepL free (500K chars/mo) is L1; Haiku-self-translate is L2 fallback. Korean and Chinese consume 1.5–3× more tokens than English. Letting Haiku translate-and-classify in one shot entangles two failure modes — strictly worse.
Trading-session field is mandatory for equity perps. Lighter equity perps trade 24/7 but NYSE is closed ~67% of the week. Pass trading_session ∈ {REGULAR, EXTENDED, CLOSED_INTERNATIONAL, CLOSED_WEEKEND} and have the prompt raise the thin-liquidity prior accordingly.
Eval harness: promptfoo + DeepEval + 50-example hand-labeled gold set (20 true-positive across 4 classes, 15 NO_CATALYST, 5 stale, 5 priced-in, 5 cross-language). Targets: JSON validity ≥99%, catalyst-type accuracy ≥85%, NO_CATALYST recall ≥90%, hallucination rate ≤3% via substring-match validator. Re-run full eval on every prompt change — arXiv 2601.22025 shows “improved” generic prompts can regress task-specific performance.
3. Latency and cadence per asset class
Headline-to-price-move latency reality check: 30-minute uniform cadence systematically misses the first 5–10 minutes of every fast catalyst. Earnings move equity prices in ~100 ms (Grégoire et al. 2025 JFE); EIA crude inventories move WTI in milliseconds; NFP/CPI move FX in milliseconds; CEX listings build CAAR mostly t-3 to t-1 before the official tweet (Blockchain Research Lab, 327-listing study, peak CAAR 14.7%). The engine cannot front-run any of these on 30-min polling. The realistic edge is cross-feed lag (Lighter mark-price lags source venues) and slow-bleed positioning (funding/OI shifts over hours).
| Asset class | Default cadence | Burst cadence on event |
|---|---|---|
| BTC/ETH/SOL majors | 5 min | 30s during FOMC/CPI/NFP windows |
| Mid-cap crypto (top 30 OI) | 10 min | 5 min |
| Long-tail/meme perps | 15 min | n/a — wash-trade dominated |
| Equity perps, NYSE open | 5 min | 30s burst for 15 min after each scheduled earnings |
| Equity perps, after-hours | 5 min | — (earnings concentrated here) |
| Equity perps, closed/weekend | 30 min | OI/funding triggers only |
| Commodity perps | 5 min | 30s burst at Wed 10:30 ET (EIA), 14:00 ET FOMC, OPEC release |
| FX perps | 15 min | 30s burst at 08:30 ET on NFP/CPI Fri/Tue |
| New listings (<7d on Lighter) | 5 min | for first 72h |
Hybrid architecture on Hetzner CPX22 (single asyncio process):
- APScheduler periodic loop with tier-1/2/3 cadences above
- Lighter WebSocket consumer on
market_stats:{id}andticker:{id}; trigger immediate scoring on|1m mark Δ| > 3σ(Lee-Mykland jump rule),|funding Δ| > 2σ over 24h cycles, orOI Δ > 5% in 5min - External news fan-in via single aiohttp pool emitting events into a priority queue that pre-empts the next periodic tick
- SQLite WAL or DuckDB for 30/90d rolling buffers (~200MB total for 100 tickers × 90d × 1m bars in float32)
Lighter mark-price lag is structural and exploitable. Per Lighter’s whitepaper and Fair Price Marking docs, mark = median of (1) index + EMA-8min(clamped book premium), (2) median of CEX marks, (3) impact price. The 8-minute EMA on the premium component means a Binance pump that completes in 30 seconds phases into Lighter mark over up to 8 minutes. Monitor Binance/NASDAQ/COMEX directly for tier-1 majors and tokenized equity perps; cross-feed lag alone is alpha. For BTC/ETH/SOL run a Binance aggTrade WS in parallel; for CRCL/INTC/AMD use Polygon.io free tier or yfinance WebSocket; for XAU/XAG/WTI use TradingView TVC or OANDA delayed.
Equity-perp closed-underlying behavior: when NYSE is closed, the underlying index feed degrades or freezes. Hyperliquid HIP-3’s TradeXYZ saw whale-triggered weekend liquidations precisely because mark drifted free of any underlying tether. Down-weight pop_score to 0.5× and up-weight oi_change and funding_z to 1.5× during these windows — positioning shifts are the cleanest signal when no underlying price exists.
4. Refined mover-ranking formula
The current |24h%| × 1.0 + volume_z × 0.5 + oi_change × 0.3 has four calibration problems: 24h% is lagging, volume z-score is corruptible by wash trading on memes, OI is under-weighted (it’s the highest-quality signal in a leveraged-only universe), and there’s no realized-vol normalization or funding term.
Recommended composite:
score =
1.0 · pop_score # max-window vol-normalized return
+ 0.7 · oi_velocity_z # 1h OI Δ z-score over 30d (raised from 0.3)
+ 0.5 · volume_z_robust # min(volume_z, trade_count_z), wash-damped
+ 0.4 · |funding_z| # funding-rate extreme magnitude
+ 0.3 · funding_flip_score # signed magnitude of regime flip
+ 0.3 · cross_venue_funding_spread_z # Lighter vs Binance/Hyperliquid
+ 0.2 · realized_vol_percentile_inv # compression → breakout candidate
− 0.5 · wash_penalty · volume_z # explicit wash-trade dampenerWhere:
pop_score = max(|r_1h|/σ_1h, |r_4h|/σ_4h, |r_24h|/σ_24h)over 30d windows. Captures both fast (EIA, earnings) and slow-bleed (unlock pressure) regimes without privileging either.wash_penalty = clamp(0, 1, (avg_trade_size / median_trade_size_30d) − 1)— high turnover relative to trade-count flags wash; alternatively usevolume / OIratio.- All raw % terms are vol-normalized via Lee-Mykland bipower variation or Garman-Klass (~7.4× more efficient than close-to-close).
- Funding-rate predictive power is modest at single-asset level (Presto Labs reports R² ≈ 12.5% over 7d on BTC) but is a cross-sectional ranking improver across the full leveraged universe.
- Cross-venue funding spread (Lighter vs Binance vs Hyperliquid) flags both arbitrage opportunity and impending mean-reversion; persistent structural spreads exist (Boros/Pendle reports 5.98–11.4% APR fixed-yield captures).
Per-class weight modifiers:
| Class | pop_score | oi_velocity | volume_z | funding sum |
|---|---|---|---|---|
| BTC/ETH/SOL | 1.0 | 0.7 | 0.5 | 0.7 |
| Mid-cap crypto | 1.0 | 0.8 | 0.4 | 1.0 |
| Long-tail/meme | 0.8 | 0.9 | 0.2 (heavy damp) | 1.0 |
| Equity perps (open) | 1.2 | 0.5 | 0.5 | 0.4 |
| Equity perps (closed) | 0.5 | 1.0 | 0.5 | 1.2 |
| Commodity perps | 1.0 | 0.7 | 0.5 | 0.6 |
| FX perps | 0.8 | 0.6 | 0.4 | 0.4 |
Cold start (<30d Lighter listing): three layers in sequence. (1) Borrow sister-cohort distribution for first 7 days — mid-cap cohort for new alts, equity cohort for new equity perps, never BTC’s distribution (too tight). (2) Empirical-Bayes shrinkage days 7–30, factor n / (n + n₀) with n₀ ≈ 7. (3) Absolute threshold backstop regardless of statistical state: |1h%| > 5% for tier-1, > 10% for long-tail. Without (3), every new listing flags as a 6σ event.
5. Alert suppression rule chain (ordered, top-down, first match wins)
| # | Rule | Action |
|---|---|---|
| 1 | Hard 4h dedup: same (ticker, catalyst_class) | DROP |
| 2 | Magnitude-escalation override (must precede dedup): |move_now| ≥ 1.5× last alert OR crossed fresh 5/10/20/50% boundary | RE-ALERT, tag escalation |
| 3 | Semantic dedup: cosine sim ≥ 0.82 (MiniLM-L6 or BGE-small) vs any 24h alert on same ticker | DROP |
| 4 | Catalyst-aware re-alert: rule 3 fires but catalyst_class differs (earnings → 8-K M&A) | RE-ALERT, tag new_catalyst |
| 5 | Sector-day clustering: ≥10 tickers same sector tag, mean ρ > 0.6, same-sign >5% in 4h | one composite alert + suppress members |
| 6 | BTC-beta filter: require |alpha_z| ≥ 2.0 AND |r_alpha| ≥ 3% | DROP if pure beta (crypto only) |
| 7 | Surprise-z gate for scheduled events: surprise_z = realized_z − expected_z < 1.0 | DROP (priced in) |
| 8 | Adaptive throttle: rolling-quantile threshold rising as daily budget depletes | DROP if score below dynamic threshold |
| 9 | Default | EMIT |
BTC-beta correction is the highest-leverage suppression rule. Rolling 30d β on 1h bars (not 90d — crypto factor structure shifts with narrative cycles), per-class benchmark factor:
| Class | Primary factor | Secondary |
|---|---|---|
| ETH, alt L1/L2 | BTC | ETH for risk-on alts |
| DeFi/AI/meme | ETH | BTC |
| Tokenized equity perps | SPY (or sector ETF: SOXX semis, KRE fintech) | DXY |
| Crypto-equity hybrids (COIN, MSTR, MARA) | BITQ + SPY | — |
| Commodity perps | DXY (inverse) | XLE |
| XAU/XAG | DXY (inverse) | TIPS 10y real yield |
| FX | DXY | regional carry pair |
Adaptive throttling with anti-starvation reservoir:
score_threshold_t = quantile_dynamic(score_history_30d,
1 − expected_remaining / hours_left)Threshold rises monotonically as budget depletes. Critical addition: maintain a rolling top-3 buffer of unsent candidates. At hour 23, if buffer contains a candidate with score > median(sent_today), force-send over budget by 1. Net effect: never lock out the actual best mover.
Daily morning brief (07:00 UTC, Telegram pinned) is recommended. Composition: overnight movers table (top-5 by alpha_z), today’s calendar (earnings/CPI/FOMC/NFP/OPEC/large unlocks/OPEX/ETF rebalance), regime tape (BTC, ETH, SPY, DXY, gold 24h + BTC dominance + aggregate funding skew), Polymarket signals (any market with ≥10pp 24h prob change on financially-relevant event), open active alpha alerts, and suppression report for transparency (“12 alerts suppressed today: 8 BTC-beta, 3 sector-day-AI, 1 dedup”).
6. Canonical catalyst taxonomy
12 classes mapped to leveraged asset classes with empirical magnitude anchors and LLM alert_priority weight (1–10).
| # | Catalyst | Magnitude (median CAR or short-window) | Horizon | Asset classes | Weight |
|---|---|---|---|---|---|
| 1 | Earnings beat/miss/guidance | Top-decile SUE: |CAR| 3–5% [-1,+1]; tail 8–15% intraday; PEAD 60d at ~1–2%/mo | 1d primary, 60d drift | Equity | 8 (+1 if surprise_z>2) |
| 2 | Regulatory (SEC, CFTC, FDA, central bank) | Crypto ETF approval +10–25%; SEC enforcement −10 to −40%; FDA AdComm ±20–60% | Hours-1d | All | 9 |
| 3 | M&A / partnership / integration | Target [-1,+1]: 15–25%; acquirer 0% to +5–7% if unanticipated | 1d | Equity, crypto | 8 |
| 4 | Insider / whale activity | Form 4 microcap open-buy: 12mo CAR 6–7.4%; whale-to-exchange transfer: 6–24h impact, ~15–25% intraday vol attribution | Form 4: weeks; whale: hours | Equity (Form 4), crypto (on-chain) | 7 / 6 |
| 5 | Macro release (CPI, NFP, FOMC, PCE, PPI) | Pre-FOMC drift +49bps (Lucca-Moench 2015 NY Fed SR512); 100bps CPI surprise → 5y UST +10bps, S&P −1 to −2% | 1d | Commodity, FX, crypto majors, equity | 9 top-tier; 7 PPI/retail |
| 6 | Geopolitical (war, sanctions, election) | OPEC quota intraday Brent ±2–6%; Russia-Ukraine onset +$19/bbl | Hours-weeks | Commodity, XAU, FX, BTC haven | 9 surprise; 5 priced-in |
| 7 | Sector rotation / narrative | Per-token magnitude noise; basket +20–50% in 1–2wk | Days-weeks | Crypto | 5 (composite alert only) |
| 8a | Token unlock | Keyrock 16k events: >1% supply −0.3% pre/post-week; >10% supply −5 to −15% in 24h | Days | Crypto | 6 if >2% supply, 4 else |
| 8b | CEX listing | 2024 reverted to +2.78% day-1 (vs +41% historic); CAAR mostly built pre-event | Hours | Crypto | 6 top-tier exchange, 3 tier-2 |
| 8c | Upgrade / mainnet | Highly variable; Merge-class 5–15% pre-event run-up | Days | Crypto | 5 |
| 8d | Hack / exploit / depeg | Host token −20 to −50% intraday; USDC 2023 depeg −12% trough; DAO crime mean −14% (arXiv 2510.00669, n=22 small) | Min-days | Crypto | 10 (always alert) |
| 9 | Liquidity event (squeeze, cascade) | Oct 11 2025: $19.2B liquidations 24h, BTC −12%; baseline BitMEX $10–20M/day | Min-hours | All leveraged | 9 if |Δfunding|>0.1%/8h or OI drops >15% in <4h |
| 10 | Technical breakout | 5–10% follow-through median on confirmed multi-day range break (crypto) | Days | All | 3 |
| 11 | ETF flow surprise | IBIT day-1 +28% AUM impact; >$500M single-day net flow → BTC ±2–4% 1–2d ahead | 1–3d | Crypto + equity | 7 if >$500M, 5 else |
| 12 | Option-related (gamma, OPEX) | Pinning 5–10d pre-OPEX at high-OI strikes; gamma squeeze 20–100% in days for low-float | Hours-days | Equity primarily, BTC options emerging | 6 UOA + price coincidence, 4 OPEX week ambient |
Tuning rule: if magnitude_observed > 2× typical, multiply weight by 1.3; if surprise_z > 2, add +2 absolute.
7. Top 5 failure modes with suppression logic
| # | Failure | Detection signal | Suppression rule |
|---|---|---|---|
| 1 | Hallucinated catalyst (ECLIPSE benchmark: Haiku-class hallucination AUC 0.59 without logprobs vs 0.89 GPT-3.5) | evidence_quotes empty OR substring-not-found in <news>; OR confidence_in_attribution < 0.4 | Force catalyst_type=NO_CATALYST, alert_priority=SUPPRESS. Validator step: all(q in news_blob for q in evidence_quotes). |
| 2 | Stale news cited as current cause (Tetlock 2011: markets systematically overreact to recombined old info) | published_at of cited evidence > 24h before move_start_utc | Set stale_news_flag=true, force is_priced_in=true, downgrade priority ≥1 step. Maintain 30d rolling embedding index to detect republished content. |
| 3 | Pump-and-dump misclassified as legitimate (Chainalysis: 24.4% of Ethereum tokens show P&D-style liquidity removal patterns) | volume_z > 5σ + market cap < $50M + thin order book + no Tier-1 news + Telegram coordinated mention spike | Pre-filter: don’t pass to Haiku at all if 24h volume < $5M. If passed: thin_liquidity_flag=true, catalyst_type=NO_CATALYST. |
| 4 | Equity-perp weekend / closed-underlying thin liquidity (5–10× lower depth in extended hours per Schwab; weekend gap volatility positively correlated with gap size) | trading_session ∈ {CLOSED_WEEKEND, CLOSED_INTERNATIONAL, EXTENDED} + perp move > 2× implied futures move | Require direct news within 6h of move_start; else thin_liquidity_flag, downgrade. Gate: emit P0/P1 on equity perps only if NYSE open OR breaking 8-K/wire tag present. |
| 5 | Cross-language miss (Korean Upbit listings empirically produce 8–62% intraday spikes — TAO/CFG/SKR/ESP cases; English-only feeds frequently miss the announcement window) | Move on Korean-domiciled token / Asian-hours move (06:00–09:00 UTC) with NO_CATALYST in EN feed | Mandatory Upbit/Bithumb/Binance.kr official-notice RSS + Weibo/WeChat crypto feeds; query Korean sources before NO_CATALYST classification. |
Tracked but lower priority: prompt injection via news content (OWASP LLM01:2025; sanitize “ignore previous” / “system:” strings, wrap all news in delimited XML and instruct model to treat content as data not instructions); JSON schema breakage (eliminated by Anthropic Structured Outputs); article republication date errors (use earliest-occurrence date, not publisher-stated); source credibility blindness (pass credibility_tier 1–4 explicitly — LLMs cannot reliably distinguish Bloomberg from a Substack).
8. Three findings that contradict the v1 spec
Finding 1: The 30-minute uniform cadence is wrong, but not in the direction the spec implies. The fix is not “scan faster across the board” — it’s tiered cadence + event-driven preemption. Earnings move equity perps in ~100 ms; EIA crude inventories move WTI in milliseconds; CEX listing CAAR builds before the announcement tweet. No periodic poller catches these. Use 5-min for tier-1 majors, 30-sec bursts during scheduled macro windows, and Lighter WebSocket triggers (jump rule, OI Δ > 5% in 5min, funding Δ > 2σ) to preempt the periodic loop. For long-tail meme perps, keep 15-min — wash-trade noise dominates faster windows. The architecture fits comfortably on a single CPX22 in one async Python process.
Finding 2: A single Haiku call per ticker is suboptimal both on cost and on accuracy. Replace it with a two-stage pipeline: BGE-base-en-v1.5 embedding prefilter + spaCy NER + FinBERT direction-validator + Haiku for natural-language rationale only. Three things this fixes simultaneously: (a) cuts Haiku tokens 40–60% by deduping and ranking news before submission, (b) provides a hallucination check Haiku cannot do alone (FinBERT direction disagreeing with Haiku direction = flag), (c) enables sub-$10/mo cost at higher recall. FinBERT/FinGPT can also replace Haiku entirely for the structured-enum classification step (FinGPT v3.3 weighted F1 0.882 on FPB, beats GPT-4); reserve the LLM call for the WIIM-style 1-sentence attribution. This is the cost-correct architecture.
Finding 3: The ranking formula’s emphasis on |24h%| privileges a lagging metric and underweights the universe’s best free signal. 24h% is by definition the slowest detector — by the time it reads 5%, the move is half-over and likely reverting. Replace with vol-normalized max-window pop (max of 1h/4h/24h, each over its own 30d σ). Raise OI weight from 0.3 to 0.7 — in a leveraged-only universe, OI velocity is the cleanest positioning signal and is structurally untouchable by wash trading. Add funding-rate z-score, funding flip, and cross-venue funding spread as separate terms — these are universal across crypto/equity/commodity/FX perps and free on Coinalyze (40/min). Add a wash-penalty subtractive term (avg_trade_size / median_trade_size_30d) so meme-perp volume z-scores don’t blow up the ranking. Add BTC-beta residualization as a hard alert gate on crypto perps: require |alpha_z| ≥ 2.0 AND |r_alpha| ≥ 3%. On a 5% BTC day, naive ranking lights up every alt; alpha-z gating surfaces only the genuinely idiosyncratic movers.
Bonus (fourth) finding worth noting: Lighter’s mark price has a structural 8-minute EMA on the premium component (per the whitepaper). A Binance pump that completes in 30 seconds phases into Lighter mark over up to 8 minutes. Monitor source venues directly (Binance WS for crypto, Polygon.io free for equities, OANDA/TVC for commodities) — the cross-feed lag is itself the alpha.
9. Build effort revision
The 92-hour estimate does not hold. The optimization surface added by the findings above adds ~35–50 hours, but materially increases the engine’s hit rate against the ≥75% catalyst-attribution and ≥80% mover-recall targets.
| Component | Original (h) | Revised (h) | Driver |
|---|---|---|---|
| Lighter SDK + universe filter (max_leverage > 1) | 8 | 8 | Unchanged |
| Periodic scanner | 12 | 8 | Simpler (just APScheduler) |
| Hybrid event-driven WS layer | 0 | 14 | New: Lighter WS, Binance WS, jump-rule triggers, priority queue |
| Free-source ingestion (RSS pool, Coinalyze, EDGAR, EIA, Polymarket) | 16 | 16 | Unchanged |
| Embedding + NER prefilter pipeline (BGE + spaCy) | 0 | 10 | New |
| FinBERT direction-validator sidecar | 0 | 6 | New |
| DeepL translation layer (KO/ZH) | 0 | 4 | New |
| Haiku integration (Structured Outputs + 3-shot) | 12 | 10 | Lighter call, structured outputs reduces retry logic |
| Source-tier table + curated wallet labels | 0 | 8 | New, manual curation |
| Mover-ranking formula (vol-normalize, funding terms, wash dampener) | 8 | 14 | Expanded |
| BTC-beta residualization (qlib rolling-window) | 0 | 8 | New |
| Suppression chain (8 ordered rules, semantic dedup, sector clustering) | 8 | 14 | Expanded |
| Adaptive throttling + reservoir | 0 | 4 | New |
| 50-example hand-labeled gold set + promptfoo + DeepEval | 0 | 12 | New, ~6h labeling + 6h harness |
| Telegram push + daily brief | 8 | 10 | Brief is new |
| Monitoring, logging, deployment | 8 | 8 | Unchanged |
| Buffer / iteration | 12 | 12 | Unchanged |
| Total | 92 | ~166 (or ~125–140 with parallelization) |
Recommendation: target 130 hours. Drop the FinBERT validator and DeepL layer to MVP-2 if 92 must be held — but the embedding prefilter, BTC-beta gate, hybrid event-driven WS, and gold-set eval harness are non-negotiable for hitting the stated accuracy targets.
Conclusion
The v1 spec’s three structural problems — uniform 30-min cadence, single-Haiku-per-ticker, and 24h%-led ranking — each cost roughly 10–20 percentage points on the catalyst-attribution and mover-recall targets, and they compound. Fixing them brings the engine into hit-rate compliance and stays within the $10/mo budget by restructuring the LLM layer rather than spending more on it. The leveraged-only universe filter is the spec’s single best decision: it makes funding rate, OI velocity, liquidation cascades, and BTC-beta residualization universal across crypto, equity, commodity, and FX perps — exploit that aggressively in the ranking formula. The free-data picture in 2026 is harsher than v1 assumed (Reuters/AP RSS dead, AlphaVantage at 25/day, no free analyst-rating-change feed, no free options unusual activity feed, Whale Alert live-API paywalled) — Coinalyze, EDGAR, GDELT, Farside, and Polymarket are the load-bearing free sources, and Polymarket’s prediction-market deltas are the most underrated leading-indicator catalyst signal available for free. Build the source-tier table by hand and trust the LLM for nothing the substring-validator cannot verify.