| name | market-scanner |
| description | Technical scan of the BET-Plus universe — computes RSI, moving averages, volume ratios, breakout levels, and trend direction from real price data, then ranks setups for entries and exits. Run this every morning after macro-analyst and bvb-news so the synthesis step has a concrete list of candidates. Uses Yahoo Finance OHLCV as the primary data source. Trigger whenever a technical scan is needed, or when evaluating a specific name's setup quality. |
Market Scanner
Quantitative scan over the BET-Plus universe. Fetches OHLCV, computes indicators, grades setups. No guesswork on current prices — always pulled fresh from Yahoo.
Universe (BET-Plus constituents)
Organized in two tiers to match the BVB index structure (~37 names in BET-Plus; we cover ~39, with a couple of legacy tickers kept for historical continuity):
Tier A — BET core (20 names, main index, high liquidity):
TLV, SNP, SNG, H2O, TGN, BRD, DIGI, EL, M, SNN, TEL, PE, FP, ONE, AQ, TRP, TTS, ATB, SFG, CFH
Where:
- PE — Premier Energy (Utilities). IPO'd on BVB 2024.
- TTS — Transport Trade Services (Industrial / logistics / Constanta port).
- CFH — Cris-Tim Family Holding (Consumer / food, entrepreneurial). IPO'd Nov 2025.
- ATB — Antibiotice (Healthcare). Promoted from extended tier to BET core to match current BET composition.
Tier B — BET-Plus beyond BET (decent liquidity, thinner names):
WINE, COTE, BVB, ROCE, ALR, BIO, CMP, IMP, LION, OIL, PPL, RRC, SIF1, SIF3, SIF5, STZ, TRANSI, UCM, EVER
Skip any symbol where 20-day average daily value traded (ADV) < 50,000 RON (PROJECT.md rule). Whenever BVB announces a BET index reshuffle, cross-check this list against the updated composition at https://bvb.ro/FinancialInstruments/Indices/IndicesProfiles.aspx?i=BET-PLUS and propose a diff in the morning briefing.
Price Data & Indicators
Call the committed script rather than regenerating the math:
node scripts/indicators.mjs --format=json TLV SNP SNG H2O TGN BRD DIGI EL M SNN TEL PE FP ONE AQ TRP TTS ATB SFG CFH
(Scan the BET core first. Run Tier B only if the core produces fewer than 3 A-grade setups or if a Tier B name has a specific catalyst today.)
Returns a JSON array, one object per symbol, with:
price, daily_change_pctrsi14, sma20, sma50, sma200volume_today, volume_avg20, volume_ratiohigh_20d, low_20d, high_52w, low_52watr14_pcttrend — one of up | down | rangen_bars — how many daily bars were available
The script fetches from Yahoo Finance (query1.finance.yahoo.com/v8/finance/chart/<SYMBOL>.RO) and computes all indicators in Python using the stdlib. Any per-symbol fetch failures are reported on stderr with [warn]; successful symbols still appear in the JSON. Exit code is 1 if any symbol was missing data, 0 if all succeeded.
Indicator definitions (for reference / audit — the script is the source of truth):
- RSI(14) — Wilder's original (no smoothing)
- SMA(n) — simple mean of last n closes
- ATR(14)% — ATR / current close × 100
- Volume ratio — today's volume / 20-day average
- Trend:
up if close > SMA50 and (SMA20 ≥ SMA50); down if close < SMA50; else range
Setup Grading
A-grade (highest conviction — ≥1 match = flag)
- Breakout: close > 20-day high AND volume ratio > 2.0 AND trend = uptrend
- Pullback to trend: trend = uptrend AND close within 2% of SMA50 AND RSI between 40-55 AND volume ratio ≥ 0.8
- Oversold within uptrend: trend = uptrend AND RSI(14) < 30 AND close > SMA200 — rare, strong when present
B-grade (good but needs confirmation)
- RSI oversold (no trend filter): RSI < 30 AND last 5 days include at least one up-close — possible mean reversion
- Volume anomaly: volume ratio > 3.0 AND close up on the day — someone is accumulating
- Higher low on RSI: price lower low vs 10 days ago, RSI higher low — divergence
C-grade (watchlist only, no trade today)
- Approaching support: trend = uptrend AND price within 3% above SMA50 but not yet touching
- Coiling: 20-day range compressed to < 6% of price, volume declining — breakout pending
- Base forming after downtrend: close > SMA20 for 5+ consecutive days after a downtrend
Sell / exit signals (for current positions)
- Trailing stop hit: current price < (peak_since_entry × 0.93) for trend rides
- Hard stop hit: current price < entry_price × 0.90 (PROJECT.md default, may be overridden per LESSONS.md)
- Take profit trigger: current price > entry_price × 1.20 for swing trades unless volume ratio still > 1.5 (momentum accelerating — hand off to synthesis for decision)
- Overbought exhaustion: RSI > 75 with MACD histogram declining 3 consecutive days
- Break of trend: trend was
uptrend, today close < SMA50 on volume ratio > 1.0 — trend compromised
Position Sizing Hint
For each A or B setup, compute a suggested initial stop based on ATR:
- Initial stop = entry × (1 - max(0.10, 2 × ATR%))
This respects the PROJECT.md 10% minimum stop but widens for high-ATR names (financials, small caps) where 10% is inside daily noise.
Dividend Awareness
Cross-check the upcoming events list from bvb-news. If a holding has ex-dividend in the next 2 trading days, the expected ex-dividend drop is NOT a sell signal — flag it explicitly so the exit logic doesn't trigger on the mechanical gap.
Output Format
🔭 MARKET SCAN — [DATE]
A-GRADE SETUPS (ranked by conviction)
[SYMBOL] [setup_type] price [X] RON RSI [Y] vol [Z]x
Entry zone: [range] Stop: [price] (-X%) Target: [price] (+X%)
Notes: [1 sentence]
B-GRADE SETUPS
[same format]
WATCHLIST (C-grade, might trigger this week)
[SYMBOL]: [what needs to happen to upgrade]
EXIT ALERTS (current positions)
[SYMBOL]: [signal] — handing to risk-monitor for decision
or
(no exit signals)
SKIPPED (illiquid or no data)
[SYMBOL]: [reason]
Sector Beta Notes (from macro-analyst/references/bvb-historical-patterns.md)
Sector-specific priors the scanner should weight when grading setups. These override intuitive assumptions — several historical correlations have broken.
Energy (SNP, SNG, TGN, RRC, OIL):
- SNP-Brent correlation 0.60-0.75 (strong, linear). Brent +4% / day →
COM-1 long SNP trade (+2-3% / 1 day).
- SNG-TTF correlation broke post-OUG 27/2022: was 0.70, now ~0.30. Don't size SNG on TTF alone.
- Regulatory shock is the dominant tail risk — fiscal-ordinance leaks (
POL-1) drop energy names -5 to -15% same-day.
- SNP removed from FTSE All Cap Feb 2026 (12-month passive-flow headwind).
Banking (TLV, BRD):
- Rate-cycle sensitivity is dominant. Each 100bp NBR move → ~8-12% re-rate over 2-3 months.
- Bank-specific fiscal shocks (turnover tax) are the idiosyncratic tail: Law 296/2023 (2%) and Law 141/2025 (4%) each produced 5-10% drawdowns.
- TLV higher-beta growth; BRD more defensive with SG-parent Stoxx Banks beta ~0.6.
- In global banking stress (SVB/CS template), RO banks -7 to -10%, recover faster than Austrian cross-border banks.
Utilities (H2O, SNN, TEL, EL, TGN, COTE, TRANSI, PE):
- H2O: 100% payout, dividend proxy; hydrology-sensitive (
COM-3 drought short). ~15.7% of BET weight.
- TEL: Jan 1 ANRE tariff reset; June 2024 +40% system services hike → +7.35% single day.
- TGN: Oct ANRE reset; +233% YoY to April 2026 on post-cap regulatory reset + BRUA + Neptun Deep offtake — mean-reversion risk.
- All: sensitive to ROBOR/NBR rates (discount rate for RAB) and state dividend directives.
Real Estate (ONE, IMP):
- Highest rate-beta on BVB. Each +100bp NBR hike → ~-15 to -20% on ONE.
- May 2026 NBR first cut is a high-conviction ONE long setup (theme: "NBR first-cut pivot").
- Catalysts: quarterly presales (>20% YoY surprise = +5-7% in 48h).
Consumer (SFG, AQ, WINE, M, CFH):
- Heterogeneous. SFG highest COVID/lockdown beta. AQ yield floor (3.4-7.7%). WINE export/FX beta.
- Mid-caps have the highest political-event beta (MedLife, Aquila, TTS, Antibiotice, Purcari all -3 to -7.6% on 2024 election shock days).
Industrial (TRP, CMP, ALR, TTS):
- ALR-LME aluminum beta only 0.24 (dampened by Hidroelectrica long-term electricity contract). Strongest at LME >$3,000/t. Inversely sensitive to RO wholesale electricity prices.
- TRP: PVC + construction PMI play.
- CMP: German auto supplier — weak on European auto slowdown.
Tech/Telecom (DIGI):
- No longer a Romanian telecom proxy — driven by Spanish M&A since 2023 (MásOrange, FTTH sale to Macquarie). +560% from May 2019 ATL to Feb 2026 ATH.
- Leverage watch: net debt/EBITDA >3.5x triggers derating risk.
BVB Operational Rules
- BVB open: 10:00 EET, close: 17:45 EET. Pre-open auction 09:45-10:00.
- Yahoo data may lag 15-20 min during RTH. Use for end-of-day analysis, not live tick decisions.
- Daily price variation limit: ±15% (tunnel). Symbols at the limit are suspended — don't attempt to trade.
- Many BVB names trade in "waves" — weeks of flat, then a 10-15% move in days. Breakout signals are rarer than on US names but often cleaner.
- January-April sees dividend-driven flows. Ex-dividend drops are mechanical; don't confuse with sell signals.
Failure Handling
- Yahoo returns no data for a symbol → try Stooq fallback
https://stooq.com/q/d/l/?s=<symbol>.ro&i=d&d1=<YYYYMMDD_start>&d2=<YYYYMMDD_end> (CSV format)
- Both fail → log the symbol as unavailable for this run. Do not fabricate a signal.
- Entire universe data fails → output a "scanner degraded" banner; synthesis should not initiate new trades without price data.
Caching
Same pattern as news/macro — in-memory cache for the run only. Do not persist between runs; the next routine fires hours later and needs fresh data.
Graphiti Sync — Skipped Setups (the blind-spot fix)
The scanner produces A-grade and B-grade candidates. The synthesis step picks which to act on. Everything that gets considered but not acted on must be persisted to graphiti — otherwise the engine has no record of the decisions it didn't make, and counterfactual learning is impossible ("a month ago we decided not to act, then the stock went up 20%").
What counts as a "skipped setup"
After synthesis runs and produces the day's actual trade decisions, the scanner emits a SkippedSetup record for every A-grade or B-grade candidate that did NOT result in a fill, with the reason for the skip. This is a separate step that runs after trade-executor, when the actual-vs-considered diff is known.
Categories of skip reason (use exactly these strings — they're the grouping keys for retrospective queries):
competing_setup_higher_conviction — another candidate ranked higher and consumed the cashliquidity_below_threshold — ADV < 50,000 RONregime_risk_off — REGIME-1 fired, cash floor hitcash_ceiling_hit — REGIME-2 inverse: already at deployment limitsector_cap_reached — 60% single-sector cap would be breachedposition_cap_reached — 30% single-stock cap would be breacheddaily_deploy_cap_hit — 50% of available cash already deployed todaychased_price — price moved >3% from signal before we could enter (the "never chase" rule)thesis_too_weak — synthesis judged conviction <5 despite scanner gradepending_event_too_close — earnings/AGM <24h away and synthesis preferred to wait for outcomeother — free-text reason in reason_detail
Ingest format
For each skipped A/B-grade candidate:
mcp__graphiti__add_memory(
name: "SkippedSetup <ticker> <date>",
episode_body: JSON.stringify({
date: "2026-04-19",
ticker: "ONE",
setup_grade: "A",
setup_type: "breakout",
score: 7.2,
price_at_skip: 12.40,
rsi14: 62,
volume_ratio: 2.3,
reason: "competing_setup_higher_conviction",
reason_detail: "TLV ranked 8.4; cash only sufficient for one position",
thesis_at_skip: "Breakout above 12.30 on 2.3x volume; NBR first-cut theme",
invalidation_window_days: 30
}),
source: "json",
source_description: "Market scanner skipped setup",
group_id: "auto_trader",
reference_time: <ISO timestamp of scan>
)
invalidation_window_days is how long we consider the skip "live" for counterfactual purposes. Defaults:
- breakout / pullback setups: 14 days
- event-driven (earnings/AGM): until the dated event + 7 days
- trend rides: 30 days
- generic: 30 days
What does NOT get a SkippedSetup
- C-grade watchlist names — they were never serious candidates today, no decision was made
- Symbols filtered out before scoring (e.g., suspended, halted, no data) — these go in the existing
SKIPPED (illiquid or no data) section, not graphiti
- Setups taken and acted on — those become
Trade nodes via trade-journal
Counterfactual closure
A separate daily routine (counterfactual-mapper) walks open SkippedSetup nodes whose invalidation_window has elapsed, fetches the price now, and attaches a Counterfactual edge with the price move since skip. This skill does NOT do that — it only emits the original record.
Failure handling
If a graphiti write fails, log to stderr + Telegram and continue. The scanner output to the synthesis step is never blocked.
Skipped setups are not mirrored to bt-gateway — they exist only in graphiti. An outage during a scan loses that day's skip records. This is acceptable because:
- A "skipped setup" represents a moment-in-time decision; the next scanner run will surface the same setup again if it's still valid, producing a new record
- Counterfactual learning loses one day of evidence, not a long-term decision audit
- Trade decisions themselves are unaffected — those still go through trade-journal → Firestore
