Cisco Time Series Model Technical Report
This work addresses time series forecasting, particularly for observability domains, but is incremental as it builds on an existing model (TimesFM) with architectural modifications.
The authors tackled the problem of time series forecasting by developing a univariate zero-shot forecaster called the Cisco Time Series Model, which achieves superior performance on observability datasets while maintaining similar performance on a general-purpose benchmark, with training on over 300B data points.
We introduce the Cisco Time Series Model, a univariate zero-shot forecaster. This time series foundation model is the result of a general architectural innovation to a time series model enabling it to accept multiresolution input, applied to a popular decoder-only time series model (TimesFM). The resulting multiresolution decoder-only model is trained on over 300B unique data points, with more than half coming from the observability domain. Quantitative and qualitative evaluations demonstrate that the resulting model achieves superior performance on observability datasets while retaining very similar performance on a standard general-purpose forecasting benchmark (GIFT-Eval), and suggest that the multiresolution structure enables the model to make more accurate predictions on long context input.