Zero‑Shot Forecasting Artifacts

Contents at a glance

• Forecast plots: Forecast Plots/
• CSV summaries: CSV Files/

Series & Windows

• Prometheus memory: 10s, 5m, 10m windows
• OpenSearch CPU: 10s, 5m, 10m windows

Forecast Horizons

• 64, 128, 256, 336 steps

Baselines & Models

• Baselines: "t" (current value) and "t-1" (last value)
• Chronos: default quantiles 0.1–0.9 to produce fan charts
• Toto: varied sampling; (32, 4) highlighted as speed/quality sweet spot

Metrics

• MASE: point-forecast accuracy versus naïve
• CRPS: calibrated uncertainty quality for predictive distributions

Evaluation Protocol

• Rolling forecast origin
• Fixed final test window
• Latency recorded per horizon for each method

Data & File Structure

• Series naming: <metric>_<window>_<source>
  • Examples: mem_util_5m_prometheus, cpu_util_10s_opensearch
• CSV files: <prediction_length>_<data_used>.csv
  • Example: 512_mem_util_5m_prometheus.csv
• Plot folders: <prediction_length>_<data_used>/ containing two plots per series
  • Files: one "chronos" plot and one "toto" plot
• Root paths in this repo:
  • Plots under Forecast Plots/
  • CSVs under CSV Files/

CSV Header Dictionary

• Toto Time: Toto inference time for the horizon (ms or s, as exported)
• Chronos Time: Chronos inference time for the horizon
• Toto MASE: MASE for Toto’s point forecast vs t baseline
• Chronos MASE: MASE for Chronos’s point forecast vs t baseline
• Toto CRPS: CRPS for Toto’s predictive distribution
• Chronos CRPS: CRPS for Chronos’s predictive distribution
• Input Length: Context length used for inference

---

Notes

• Time units in exported CSVs may be milliseconds or seconds depending on source tool.
• Chronos fan charts are rendered from the 0.1–0.9 quantiles.
• Toto results reflect different sampling configs; (32, 4) is a good trade‑off.

Baselines Per Horizon

• Definition: Let the last observed value at the forecast origin be y_t, and the previous value be y_{t-1}. For a horizon H (e.g., 64/128/256/336), the baselines are:
  • "t" baseline (persistence): for all h = 1..H, ŷ_{t+h} = y_t.
  • "t-1" baseline: for all h = 1..H, ŷ_{t+h} = y_{t-1}.
• Usage: These baselines produce a full H-step forecast vector by repeating a single observed value across all steps.
• Metrics: MASE and CRPS are computed per horizon window using these baseline definitions for comparison.
• Edge cases: If y_{t-1} is not available (very early in the series), the "t-1" baseline is skipped or marked N/A for that origin.