Phase 1 · Event Encoder (20 channels)

Priority: High Status: Pending Depends on: None

Context

• Brainstorm report: plans/reports/brainstorm-260423-0059-event-conditioned-kronos-research.md
• Economic calendar data: economic_calendar table (Supabase)
• Event types: FOMC, CPI, NFP, GDP, PCE, earnings, rate decisions

Overview

Build a standalone module that maps economic calendar events + cross-asset leaders to aligned (T, 20) feature tensors matching OHLCV candle timestamps.

Requirements

Functional

• Query economic_calendar for events within a date range
• Compute continuous surprise z-score per event type (NOT sign() — magnitude matters)
• Align events to OHLCV timestamps (daily: fire on event day; intraday: fire on release hour)
• Encode days-until-event as sinusoidal features (anticipatory vol)
• Fetch cross-asset leader returns (BTC, SPY, DXY, VIX) with prior-close-only enforcement
• Output: numpy array (seq_len, 20) aligned with OHLCV input

Non-Functional

• No external API calls at inference time (calendar + leader data pre-loaded)
• Support both daily and intraday timeframes
• Cache historical surprise std per event type
• Mask cross-asset channels when leader market closed (no forward-fill bias)

Architecture

Event Channels (20 features per timestep)

Index	Channel	Values	Source
0	is_fomc	0/1	economic_calendar
1	is_cpi	0/1	economic_calendar
2	is_nfp	0/1	economic_calendar
3	is_gdp	0/1	economic_calendar
4	is_pce	0/1	economic_calendar
5	cpi_surprise_z	continuous float	(actual - forecast) / σ, clip to [-3,3]
6	fomc_hawkish_score	continuous float	statement tone NLP, range [-1,1]
7	nfp_surprise_z	continuous float	(actual - forecast) / σ, clip to [-3,3]
8	is_earnings	0/1	earnings calendar
9	is_rate_decision	0/1	central_bank_rates
10	days_to_fomc_sin	[-1,1]	sin(2π · days_to_next_fomc / 30)
11	days_to_fomc_cos	[-1,1]	cos(2π · days_to_next_fomc / 30)
12	days_to_cpi_sin	[-1,1]	sin(2π · days_to_next_cpi / 30)
13	days_to_cpi_cos	[-1,1]	cos(2π · days_to_next_cpi / 30)
14	btc_log_return_1h	continuous	prior bar only
15	spy_log_return_1h	continuous	prior bar only, mask when NYSE closed
16	dxy_log_return_1h	continuous	prior bar only, mask when FX session closed
17	vix_level_z	continuous	z-scored VIX level
18	reserved	0	future expansion
19	reserved	0	future expansion

Surprise Computation (continuous, not discrete)

surprise_z = (actual - forecast) / rolling_std(past_20_surprises)
# keep continuous value, clip to [-3, 3] to bound outliers
# DO NOT quantize to {-1, 0, +1} — magnitude carries signal

Cross-Asset Leader — Leakage Prevention

def get_leader_return(leader_symbol, target_timestamp):
    # MUST use bar STRICTLY before target_timestamp
    prior_bar = query_ohlcv(leader_symbol, end=target_timestamp - 1_bar)
    return log(prior_bar.close / prior_prior_bar.close)
    # If market closed (SPY on weekend, during target's session):
    # return (value=0, mask=1) — paired mask channel

Leakage rule: Leader bar timestamp MUST be < target bar timestamp. Never == or >.

Intraday Alignment

For hourly candles:

• CPI/NFP release at 8:30 ET (13:30 UTC) → flag on 13:00 or 14:00 UTC bar
• FOMC statement at 14:00 ET (19:00 UTC) → flag on 19:00 or 20:00 UTC bar
• For daily candles: flag entire day

Implementation Steps

1. Create kronos-service/event_encoder.py
2. Implement EventEncoder.__init__() — load calendar, compute surprise stats
3. Implement EventEncoder.encode(timestamps, timeframe) — returns (T, 20) tensor
4. Implement EventEncoder._compute_surprise_z() — continuous z-score per event
5. Implement EventEncoder._days_until_sinusoidal() — channels 10–13
6. Implement EventEncoder._fetch_leader_returns() — BTC/SPY/DXY/VIX w/ leakage guard
7. Implement EventEncoder._align_to_candles() — timestamp alignment
8. Add unit tests for edge cases (no events, multi-event day, market-closed leader, leakage guard)

Key Files

• Create: kronos-service/event_encoder.py
• Read: scripts/kronos-batch-predict.py (DB connection pattern)
• Read: scripts/daily-refresh.py (FRED_SERIES list)
• Read: Supabase economic_calendar schema

Success Criteria

• Produces (T, 20) array for any symbol/timeframe/date_range
• Handles missing forecast/actual gracefully (outputs 0)
• Unit tests pass for: no events, single event, multi-event, intraday alignment
• Cross-asset leader leakage guard verified (timestamp strict <)
• Market-closed leader channels correctly masked (no forward-fill)
• Inference <5ms per encode call (added leader fetch cost)