Cenace s3 daily benchmark implementation

[elmartinj]

Adds an S3-ready CENACE daily data update pipeline.

• Restores CENACE raw CSV transform into unique_id/ds/y format
• Writes hourly daily parquet partitions directly, avoiding a giant processed CSV as the deployed store
• Adds CENACE_DATA_ROOT so local paths can be swapped for the Modal S3 mount
• Adds a Modal runner that writes raw ZIPs and processed parquet to s3://impermanent-benchmark/v0.1.0/cenace
• Adds a scheduled GitHub Actions workflow at 06:30 UTC / 00:30 Mexico time
• Adds beautifulsoup4 as an explicit dependency

Validation:

• uv run pre-commit run --all-files
• local CENACE pipeline smoke for 2025-01-02
• Modal/S3 smoke for 2025-01-02

[elmartinj]

Summary

Adds the daily CENACE S3 benchmark pipeline.

This PR wires CENACE data updates, hourly parquet partitioning, Modal execution, S3-mounted reads/writes, and daily forecast/evaluation runs for the CENACE benchmark.

What changed

• Added S3-ready CENACE data update pipeline.

• Added Modal runner for daily CENACE data updates.

• Restored/added CENACE transform logic for selected benchmark nodes.

• Writes hourly parquet partitions under processed-events/hourly/year=YYYY/month=MM/day=DD/.

• Added S3-safe parquet writes using local temp files before copying into the mounted bucket.

• Added Modal forecast/evaluation runner for CENACE.

• Split forecast and evaluation into separate Modal entrypoints:

  • forecast
  • evaluate

• Added Makefile targets:

  • update-cenace-data
  • update-cenace-forecast
  • update-cenace-evaluate

• Added scheduled GitHub workflow:

  • fetch today’s CENACE data
  • evaluate yesterday’s forecast when available
  • forecast the next CENACE day

Benchmark models

Initial daily benchmark runs CPU-safe models only:

• seasonal_naive
• historic_average
• auto_ets
• auto_ces
• dynamic_optimized_theta

GPU/foundation models are left for a follow-up because they require heavier runtime/cost handling.

Validation

Local checks:

uv run pre-commit run --all-files

Passed:

• ruff
• ruff-format
• mypy
• toml-sort-fix

Modal/S3 data smoke:

make update-cenace-data EXECUTION_DATE=2024-12-31
make update-cenace-data EXECUTION_DATE=2025-01-01

Successfully wrote:

• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=01/series.parquet
• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=02/series.parquet

Forecast smoke:

make update-cenace-forecast CUTOFF=2025-01-02T23:00:00

Successfully wrote forecasts for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/forecasts/hourly/<model>/year=2025/month=01/day=02/forecasts.parquet

Evaluation smoke:

make update-cenace-evaluate CUTOFF=2025-01-02T23:00:00

Successfully wrote metrics for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/evaluations/hourly/<model>/year=2025/month=01/day=02/metrics.parquet

Notes

The evaluation step in the scheduled workflow is marked continue-on-error: true so the first live run does not block tomorrow’s forecast if the previous day’s forecast does not exist yet.

[elmartinj]

Adds an S3-ready CENACE daily data update pipeline.

• Restores CENACE raw CSV transform into unique_id/ds/y format
• Writes hourly daily parquet partitions directly, avoiding a giant processed CSV as the deployed store
• Adds CENACE_DATA_ROOT so local paths can be swapped for the Modal S3 mount
• Adds a Modal runner that writes raw ZIPs and processed parquet to s3://impermanent-benchmark/v0.1.0/cenace
• Adds a scheduled GitHub Actions workflow at 06:30 UTC / 00:30 Mexico time
• Adds beautifulsoup4 as an explicit dependency

Validation:

• uv run pre-commit run --all-files
• local CENACE pipeline smoke for 2025-01-02
• Modal/S3 smoke for 2025-01-02

[elmartinj]

Summary

Adds the daily CENACE S3 benchmark pipeline.

This PR wires CENACE data updates, hourly parquet partitioning, Modal execution, S3-mounted reads/writes, and daily forecast/evaluation runs for the CENACE benchmark.

What changed

• Added S3-ready CENACE data update pipeline.

• Added Modal runner for daily CENACE data updates.

• Restored/added CENACE transform logic for selected benchmark nodes.

• Writes hourly parquet partitions under processed-events/hourly/year=YYYY/month=MM/day=DD/.

• Added S3-safe parquet writes using local temp files before copying into the mounted bucket.

• Added Modal forecast/evaluation runner for CENACE.

• Split forecast and evaluation into separate Modal entrypoints:

  • forecast
  • evaluate

• Added Makefile targets:

  • update-cenace-data
  • update-cenace-forecast
  • update-cenace-evaluate

• Added scheduled GitHub workflow:

  • fetch today’s CENACE data
  • evaluate yesterday’s forecast when available
  • forecast the next CENACE day

Benchmark models

Initial daily benchmark runs CPU-safe models only:

• seasonal_naive
• historic_average
• auto_ets
• auto_ces
• dynamic_optimized_theta

GPU/foundation models are left for a follow-up because they require heavier runtime/cost handling.

Validation

Local checks:

uv run pre-commit run --all-files

Passed:

• ruff
• ruff-format
• mypy
• toml-sort-fix

Modal/S3 data smoke:

make update-cenace-data EXECUTION_DATE=2024-12-31
make update-cenace-data EXECUTION_DATE=2025-01-01

Successfully wrote:

• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=01/series.parquet
• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=02/series.parquet

Forecast smoke:

make update-cenace-forecast CUTOFF=2025-01-02T23:00:00

Successfully wrote forecasts for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/forecasts/hourly/<model>/year=2025/month=01/day=02/forecasts.parquet

Evaluation smoke:

make update-cenace-evaluate CUTOFF=2025-01-02T23:00:00

Successfully wrote metrics for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/evaluations/hourly/<model>/year=2025/month=01/day=02/metrics.parquet

Notes

The evaluation step in the scheduled workflow is marked continue-on-error: true so the first live run does not block tomorrow’s forecast if the previous day’s forecast does not exist yet.

[elmartinj]

Adds an S3-ready CENACE daily data update pipeline.

• Restores CENACE raw CSV transform into unique_id/ds/y format
• Writes hourly daily parquet partitions directly, avoiding a giant processed CSV as the deployed store
• Adds CENACE_DATA_ROOT so local paths can be swapped for the Modal S3 mount
• Adds a Modal runner that writes raw ZIPs and processed parquet to s3://impermanent-benchmark/v0.1.0/cenace
• Adds a scheduled GitHub Actions workflow at 06:30 UTC / 00:30 Mexico time
• Adds beautifulsoup4 as an explicit dependency

Validation:

• uv run pre-commit run --all-files
• local CENACE pipeline smoke for 2025-01-02
• Modal/S3 smoke for 2025-01-02

[elmartinj]

Summary

Adds the daily CENACE S3 benchmark pipeline.

This PR wires CENACE data updates, hourly parquet partitioning, Modal execution, S3-mounted reads/writes, and daily forecast/evaluation runs for the CENACE benchmark.

What changed

• Added S3-ready CENACE data update pipeline.

• Added Modal runner for daily CENACE data updates.

• Restored/added CENACE transform logic for selected benchmark nodes.

• Writes hourly parquet partitions under processed-events/hourly/year=YYYY/month=MM/day=DD/.

• Added S3-safe parquet writes using local temp files before copying into the mounted bucket.

• Added Modal forecast/evaluation runner for CENACE.

• Split forecast and evaluation into separate Modal entrypoints:

  • forecast
  • evaluate

• Added Makefile targets:

  • update-cenace-data
  • update-cenace-forecast
  • update-cenace-evaluate

• Added scheduled GitHub workflow:

  • fetch today’s CENACE data
  • evaluate yesterday’s forecast when available
  • forecast the next CENACE day

Benchmark models

Initial daily benchmark runs CPU-safe models only:

• seasonal_naive
• historic_average
• auto_ets
• auto_ces
• dynamic_optimized_theta

GPU/foundation models are left for a follow-up because they require heavier runtime/cost handling.

Validation

Local checks:

uv run pre-commit run --all-files

Passed:

• ruff
• ruff-format
• mypy
• toml-sort-fix

Modal/S3 data smoke:

make update-cenace-data EXECUTION_DATE=2024-12-31
make update-cenace-data EXECUTION_DATE=2025-01-01

Successfully wrote:

• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=01/series.parquet
• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=02/series.parquet

Forecast smoke:

make update-cenace-forecast CUTOFF=2025-01-02T23:00:00

Successfully wrote forecasts for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/forecasts/hourly/<model>/year=2025/month=01/day=02/forecasts.parquet

Evaluation smoke:

make update-cenace-evaluate CUTOFF=2025-01-02T23:00:00

Successfully wrote metrics for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/evaluations/hourly/<model>/year=2025/month=01/day=02/metrics.parquet

Notes

The evaluation step in the scheduled workflow is marked continue-on-error: true so the first live run does not block tomorrow’s forecast if the previous day’s forecast does not exist yet.

[elmartinj]

Adds an S3-ready CENACE daily data update pipeline.

• Restores CENACE raw CSV transform into unique_id/ds/y format
• Writes hourly daily parquet partitions directly, avoiding a giant processed CSV as the deployed store
• Adds CENACE_DATA_ROOT so local paths can be swapped for the Modal S3 mount
• Adds a Modal runner that writes raw ZIPs and processed parquet to s3://impermanent-benchmark/v0.1.0/cenace
• Adds a scheduled GitHub Actions workflow at 06:30 UTC / 00:30 Mexico time
• Adds beautifulsoup4 as an explicit dependency

Validation:

• uv run pre-commit run --all-files
• local CENACE pipeline smoke for 2025-01-02
• Modal/S3 smoke for 2025-01-02

[elmartinj]

Summary

Adds the daily CENACE S3 benchmark pipeline.

This PR wires CENACE data updates, hourly parquet partitioning, Modal execution, S3-mounted reads/writes, and daily forecast/evaluation runs for the CENACE benchmark.

What changed

• Added S3-ready CENACE data update pipeline.

• Added Modal runner for daily CENACE data updates.

• Restored/added CENACE transform logic for selected benchmark nodes.

• Writes hourly parquet partitions under processed-events/hourly/year=YYYY/month=MM/day=DD/.

• Added S3-safe parquet writes using local temp files before copying into the mounted bucket.

• Added Modal forecast/evaluation runner for CENACE.

• Split forecast and evaluation into separate Modal entrypoints:

  • forecast
  • evaluate

• Added Makefile targets:

  • update-cenace-data
  • update-cenace-forecast
  • update-cenace-evaluate

• Added scheduled GitHub workflow:

  • fetch today’s CENACE data
  • evaluate yesterday’s forecast when available
  • forecast the next CENACE day

Benchmark models

Initial daily benchmark runs CPU-safe models only:

• seasonal_naive
• historic_average
• auto_ets
• auto_ces
• dynamic_optimized_theta

GPU/foundation models are left for a follow-up because they require heavier runtime/cost handling.

Validation

Local checks:

uv run pre-commit run --all-files

Passed:

• ruff
• ruff-format
• mypy
• toml-sort-fix

Modal/S3 data smoke:

make update-cenace-data EXECUTION_DATE=2024-12-31
make update-cenace-data EXECUTION_DATE=2025-01-01

Successfully wrote:

• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=01/series.parquet
• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=02/series.parquet

Forecast smoke:

make update-cenace-forecast CUTOFF=2025-01-02T23:00:00

Successfully wrote forecasts for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/forecasts/hourly/<model>/year=2025/month=01/day=02/forecasts.parquet

Evaluation smoke:

make update-cenace-evaluate CUTOFF=2025-01-02T23:00:00

Successfully wrote metrics for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/evaluations/hourly/<model>/year=2025/month=01/day=02/metrics.parquet

Notes

The evaluation step in the scheduled workflow is marked continue-on-error: true so the first live run does not block tomorrow’s forecast if the previous day’s forecast does not exist yet.

[elmartinj]

Adds an S3-ready CENACE daily data update pipeline.

• Restores CENACE raw CSV transform into unique_id/ds/y format
• Writes hourly daily parquet partitions directly, avoiding a giant processed CSV as the deployed store
• Adds CENACE_DATA_ROOT so local paths can be swapped for the Modal S3 mount
• Adds a Modal runner that writes raw ZIPs and processed parquet to s3://impermanent-benchmark/v0.1.0/cenace
• Adds a scheduled GitHub Actions workflow at 06:30 UTC / 00:30 Mexico time
• Adds beautifulsoup4 as an explicit dependency

Validation:

• uv run pre-commit run --all-files
• local CENACE pipeline smoke for 2025-01-02
• Modal/S3 smoke for 2025-01-02

[elmartinj]

Summary

Adds the daily CENACE S3 benchmark pipeline.

This PR wires CENACE data updates, hourly parquet partitioning, Modal execution, S3-mounted reads/writes, and daily forecast/evaluation runs for the CENACE benchmark.

What changed

• Added S3-ready CENACE data update pipeline.

• Added Modal runner for daily CENACE data updates.

• Restored/added CENACE transform logic for selected benchmark nodes.

• Writes hourly parquet partitions under processed-events/hourly/year=YYYY/month=MM/day=DD/.

• Added S3-safe parquet writes using local temp files before copying into the mounted bucket.

• Added Modal forecast/evaluation runner for CENACE.

• Split forecast and evaluation into separate Modal entrypoints:

  • forecast
  • evaluate

• Added Makefile targets:

  • update-cenace-data
  • update-cenace-forecast
  • update-cenace-evaluate

• Added scheduled GitHub workflow:

  • fetch today’s CENACE data
  • evaluate yesterday’s forecast when available
  • forecast the next CENACE day

Benchmark models

Initial daily benchmark runs CPU-safe models only:

• seasonal_naive
• historic_average
• auto_ets
• auto_ces
• dynamic_optimized_theta

GPU/foundation models are left for a follow-up because they require heavier runtime/cost handling.

Validation

Local checks:

uv run pre-commit run --all-files

Passed:

• ruff
• ruff-format
• mypy
• toml-sort-fix

Modal/S3 data smoke:

make update-cenace-data EXECUTION_DATE=2024-12-31
make update-cenace-data EXECUTION_DATE=2025-01-01

Successfully wrote:

• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=01/series.parquet
• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=02/series.parquet

Forecast smoke:

make update-cenace-forecast CUTOFF=2025-01-02T23:00:00

Successfully wrote forecasts for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/forecasts/hourly/<model>/year=2025/month=01/day=02/forecasts.parquet

Evaluation smoke:

make update-cenace-evaluate CUTOFF=2025-01-02T23:00:00

Successfully wrote metrics for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/evaluations/hourly/<model>/year=2025/month=01/day=02/metrics.parquet

Notes

The evaluation step in the scheduled workflow is marked continue-on-error: true so the first live run does not block tomorrow’s forecast if the previous day’s forecast does not exist yet.

[elmartinj]

Adds an S3-ready CENACE daily data update pipeline.

• Restores CENACE raw CSV transform into unique_id/ds/y format
• Writes hourly daily parquet partitions directly, avoiding a giant processed CSV as the deployed store
• Adds CENACE_DATA_ROOT so local paths can be swapped for the Modal S3 mount
• Adds a Modal runner that writes raw ZIPs and processed parquet to s3://impermanent-benchmark/v0.1.0/cenace
• Adds a scheduled GitHub Actions workflow at 06:30 UTC / 00:30 Mexico time
• Adds beautifulsoup4 as an explicit dependency

Validation:

• uv run pre-commit run --all-files
• local CENACE pipeline smoke for 2025-01-02
• Modal/S3 smoke for 2025-01-02

[elmartinj]

Summary

Adds the daily CENACE S3 benchmark pipeline.

This PR wires CENACE data updates, hourly parquet partitioning, Modal execution, S3-mounted reads/writes, and daily forecast/evaluation runs for the CENACE benchmark.

What changed

• Added S3-ready CENACE data update pipeline.

• Added Modal runner for daily CENACE data updates.

• Restored/added CENACE transform logic for selected benchmark nodes.

• Writes hourly parquet partitions under processed-events/hourly/year=YYYY/month=MM/day=DD/.

• Added S3-safe parquet writes using local temp files before copying into the mounted bucket.

• Added Modal forecast/evaluation runner for CENACE.

• Split forecast and evaluation into separate Modal entrypoints:

  • forecast
  • evaluate

• Added Makefile targets:

  • update-cenace-data
  • update-cenace-forecast
  • update-cenace-evaluate

• Added scheduled GitHub workflow:

  • fetch today’s CENACE data
  • evaluate yesterday’s forecast when available
  • forecast the next CENACE day

Benchmark models

Initial daily benchmark runs CPU-safe models only:

• seasonal_naive
• historic_average
• auto_ets
• auto_ces
• dynamic_optimized_theta

GPU/foundation models are left for a follow-up because they require heavier runtime/cost handling.

Validation

Local checks:

uv run pre-commit run --all-files

Passed:

• ruff
• ruff-format
• mypy
• toml-sort-fix

Modal/S3 data smoke:

make update-cenace-data EXECUTION_DATE=2024-12-31
make update-cenace-data EXECUTION_DATE=2025-01-01

Successfully wrote:

• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=01/series.parquet
• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=02/series.parquet

Forecast smoke:

make update-cenace-forecast CUTOFF=2025-01-02T23:00:00

Successfully wrote forecasts for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/forecasts/hourly/<model>/year=2025/month=01/day=02/forecasts.parquet

Evaluation smoke:

make update-cenace-evaluate CUTOFF=2025-01-02T23:00:00

Successfully wrote metrics for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/evaluations/hourly/<model>/year=2025/month=01/day=02/metrics.parquet

Notes

The evaluation step in the scheduled workflow is marked continue-on-error: true so the first live run does not block tomorrow’s forecast if the previous day’s forecast does not exist yet.

[elmartinj]

Adds an S3-ready CENACE daily data update pipeline.

• Restores CENACE raw CSV transform into unique_id/ds/y format
• Writes hourly daily parquet partitions directly, avoiding a giant processed CSV as the deployed store
• Adds CENACE_DATA_ROOT so local paths can be swapped for the Modal S3 mount
• Adds a Modal runner that writes raw ZIPs and processed parquet to s3://impermanent-benchmark/v0.1.0/cenace
• Adds a scheduled GitHub Actions workflow at 06:30 UTC / 00:30 Mexico time
• Adds beautifulsoup4 as an explicit dependency

Validation:

• uv run pre-commit run --all-files
• local CENACE pipeline smoke for 2025-01-02
• Modal/S3 smoke for 2025-01-02

[elmartinj]

Summary

Adds the daily CENACE S3 benchmark pipeline.

This PR wires CENACE data updates, hourly parquet partitioning, Modal execution, S3-mounted reads/writes, and daily forecast/evaluation runs for the CENACE benchmark.

What changed

• Added S3-ready CENACE data update pipeline.

• Added Modal runner for daily CENACE data updates.

• Restored/added CENACE transform logic for selected benchmark nodes.

• Writes hourly parquet partitions under processed-events/hourly/year=YYYY/month=MM/day=DD/.

• Added S3-safe parquet writes using local temp files before copying into the mounted bucket.

• Added Modal forecast/evaluation runner for CENACE.

• Split forecast and evaluation into separate Modal entrypoints:

  • forecast
  • evaluate

• Added Makefile targets:

  • update-cenace-data
  • update-cenace-forecast
  • update-cenace-evaluate

• Added scheduled GitHub workflow:

  • fetch today’s CENACE data
  • evaluate yesterday’s forecast when available
  • forecast the next CENACE day

Benchmark models

Initial daily benchmark runs CPU-safe models only:

• seasonal_naive
• historic_average
• auto_ets
• auto_ces
• dynamic_optimized_theta

GPU/foundation models are left for a follow-up because they require heavier runtime/cost handling.

Validation

Local checks:

uv run pre-commit run --all-files

Passed:

• ruff
• ruff-format
• mypy
• toml-sort-fix

Modal/S3 data smoke:

make update-cenace-data EXECUTION_DATE=2024-12-31
make update-cenace-data EXECUTION_DATE=2025-01-01

Successfully wrote:

• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=01/series.parquet
• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=02/series.parquet

Forecast smoke:

make update-cenace-forecast CUTOFF=2025-01-02T23:00:00

Successfully wrote forecasts for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/forecasts/hourly/<model>/year=2025/month=01/day=02/forecasts.parquet

Evaluation smoke:

make update-cenace-evaluate CUTOFF=2025-01-02T23:00:00

Successfully wrote metrics for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/evaluations/hourly/<model>/year=2025/month=01/day=02/metrics.parquet

Notes

The evaluation step in the scheduled workflow is marked continue-on-error: true so the first live run does not block tomorrow’s forecast if the previous day’s forecast does not exist yet.

[elmartinj]

Adds an S3-ready CENACE daily data update pipeline.

• Restores CENACE raw CSV transform into unique_id/ds/y format
• Writes hourly daily parquet partitions directly, avoiding a giant processed CSV as the deployed store
• Adds CENACE_DATA_ROOT so local paths can be swapped for the Modal S3 mount
• Adds a Modal runner that writes raw ZIPs and processed parquet to s3://impermanent-benchmark/v0.1.0/cenace
• Adds a scheduled GitHub Actions workflow at 06:30 UTC / 00:30 Mexico time
• Adds beautifulsoup4 as an explicit dependency

Validation:

• uv run pre-commit run --all-files
• local CENACE pipeline smoke for 2025-01-02
• Modal/S3 smoke for 2025-01-02

[elmartinj]

Summary

Adds the daily CENACE S3 benchmark pipeline.

This PR wires CENACE data updates, hourly parquet partitioning, Modal execution, S3-mounted reads/writes, and daily forecast/evaluation runs for the CENACE benchmark.

What changed

• Added S3-ready CENACE data update pipeline.

• Added Modal runner for daily CENACE data updates.

• Restored/added CENACE transform logic for selected benchmark nodes.

• Writes hourly parquet partitions under processed-events/hourly/year=YYYY/month=MM/day=DD/.

• Added S3-safe parquet writes using local temp files before copying into the mounted bucket.

• Added Modal forecast/evaluation runner for CENACE.

• Split forecast and evaluation into separate Modal entrypoints:

  • forecast
  • evaluate

• Added Makefile targets:

  • update-cenace-data
  • update-cenace-forecast
  • update-cenace-evaluate

• Added scheduled GitHub workflow:

  • fetch today’s CENACE data
  • evaluate yesterday’s forecast when available
  • forecast the next CENACE day

Benchmark models

Initial daily benchmark runs CPU-safe models only:

• seasonal_naive
• historic_average
• auto_ets
• auto_ces
• dynamic_optimized_theta

GPU/foundation models are left for a follow-up because they require heavier runtime/cost handling.

Validation

Local checks:

uv run pre-commit run --all-files

Passed:

• ruff
• ruff-format
• mypy
• toml-sort-fix

Modal/S3 data smoke:

make update-cenace-data EXECUTION_DATE=2024-12-31
make update-cenace-data EXECUTION_DATE=2025-01-01

Successfully wrote:

• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=01/series.parquet
• s3://impermanent-benchmark/v0.1.0/cenace/processed-events/hourly/year=2025/month=01/day=02/series.parquet

Forecast smoke:

make update-cenace-forecast CUTOFF=2025-01-02T23:00:00

Successfully wrote forecasts for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/forecasts/hourly/<model>/year=2025/month=01/day=02/forecasts.parquet

Evaluation smoke:

make update-cenace-evaluate CUTOFF=2025-01-02T23:00:00

Successfully wrote metrics for all 5 CPU models under:

s3://impermanent-benchmark/v0.1.0/cenace/evaluations/hourly/<model>/year=2025/month=01/day=02/metrics.parquet

Notes

The evaluation step in the scheduled workflow is marked continue-on-error: true so the first live run does not block tomorrow’s forecast if the previous day’s forecast does not exist yet.