| name | time-series-setup |
| description | Configure GAFIME Engine v0.4.x time-series candidate search for temporal/sequential data. Use when the user has temporal data, transaction logs, time-stamped records, or asks things like "set up time series features", "time series feature engineering", "velocity features", "rolling windows", or "regime features". |
Time-Series Setup
Guide users through configuring Engine-level time-series candidate search.
Instructions
-
Ask the user about their data:
- What is the group column? (e.g.,
customer_id, user_id, account_id)
- What is the time column? (e.g.,
date, timestamp, transaction_date)
- What is the target? (e.g.,
churn, fraud, default)
- What is the time granularity? (hourly, daily, weekly, monthly)
-
If they have the data file, run the detector:
python .claude/skills/time-series-setup/scripts/detect_time_structure.py "<file_path>"
-
Based on the data's time span and granularity, recommend window sizes:
| Granularity | Recommended Windows | Rationale |
|---|
| Hourly | [6, 12, 24, 48, 168] | 6h, 12h, 1d, 2d, 1w |
| Daily | [7, 14, 30, 60, 90, 180, 360] | 1w, 2w, 1m, 2m, 3m, 6m, 1y |
| Weekly | [4, 8, 13, 26, 52] | 1m, 2m, 1q, 6m, 1y |
| Monthly | [3, 6, 12, 24] | 1q, 6m, 1y, 2y |
-
Recommend whether to enable calculus features:
- Velocity (1st derivative): Always useful — captures rate of change
- Acceleration (2nd derivative): Useful for detecting trend reversals
- Momentum (jerk): Useful for volatile financial data
- Integral: Useful for cumulative metrics
- Volatility ratios: Great for financial data
- Trend strength: Good for detecting regime changes
-
Generate the complete setup code:
from gafime import ComputeBudget, EngineConfig, GafimeEngine
config = EngineConfig(
backend="auto",
metric_names=("pearson", "r2"),
enable_time_series_functions=True,
time_series_lags=(1, 2, 4, 8, 16),
time_series_windows=(4, 8, 16, 32),
budget=ComputeBudget(max_time_series_candidates=100_000),
)
report = GafimeEngine(config).analyze(X_train, y_train, feature_names)
-
Warn about common pitfalls:
- Windows longer than the data span → produces all NaN
- Too many windows × features → combinatorial explosion
- Missing timestamps → Polars will error on sort
- Non-numeric feature columns → need encoding first
- If v0.4.x discrete functions are enabled after time-series aggregation,
thresholds must be fit only on training folds.
-
If the user wants threshold or regime-style time-series signals, recommend
running GafimeEngine with enable_discrete_functions=True after
aggregation. On GPU, use discrete_mode="soft"; hard mode is C++ Core only.
Example
User says: "I have daily transaction data per customer, need to predict churn"
Result: "For daily data, I recommend windows [7, 14, 30, 60, 90, 180, 360]. Enable calculus features — velocity and acceleration are great for detecting customers who are slowing down before churning. Here's your setup code: ..."
