Event-Driven Signals¶
Trading on sparse events like earnings.
Handling Sparse Data¶
Event data has values only on event days, NaN otherwise:
AAPL MSFT
2024-01-15 NaN NaN
2024-01-16 0.05 NaN AAPL beat by 5%
2024-01-17 NaN NaN
2024-01-18 NaN -0.02 MSFT missed by 2%
Key Operations¶
Forward Fill¶
Hold a signal for N days after the event:
Check Validity¶
Did an event occur?
Event-Based Rolling¶
Roll over N events (not N days):
ts_mean_events(field('earnings_surprise'), 4) # Mean of last 4 quarters
ts_std_events(field('earnings_surprise'), 8) # Std of last 8 quarters
PEAD Signal¶
Post-Earnings Announcement Drift:
# Basic: go long positive surprises
let surprise = field('earnings_surprise'),
held = fill_forward(surprise, 60)
in rank(held) - 0.5
SUE (Standardized Unexpected Earnings)¶
Academic version normalizes by historical surprise volatility:
let surprise = field('earnings_surprise'),
surprise_std = ts_std_events(surprise, 8),
sue = surprise / surprise_std,
held = fill_forward(sue, 60)
in rank(held) - 0.5
Earnings Momentum¶
Trend in earnings surprises:
let surprise = field('earnings_surprise'),
avg_surprise = ts_mean_events(surprise, 4),
held = fill_forward(avg_surprise, 60)
in rank(held) - 0.5
Announcement Returns¶
Trade on the announcement day return:
let announce = is_valid(field('earnings_surprise')),
ann_ret = where(announce, returns(1), None),
held = fill_forward(ann_ret, 60)
in rank(-held) - 0.5
This fades the immediate reaction.
Event Count Filter¶
Only trade stocks with enough history:
let surprise = field('earnings_surprise'),
count = ts_count_events(surprise, 252),
sue = surprise / ts_std_events(surprise, 8),
held = fill_forward(sue, 60)
in where(count >= 4, rank(held) - 0.5, None)
Multiple Event Types¶
Combine different events: