This study investigates the potential of TimeGPT, a pretrained foundation model for timeseries, for forecasting daily emergency call volumes. Real-world emergency call data, comprising 2,101 daily observations, are used to evaluate both univariate and multivariate forecasting settings under one-step-ahead and seven-day multi-step-ahead scenarios. The multivariate forecasting incorporates temporal, meteorological, demographic, and socioeconomic variables. A chronological training, validation, and test split is adopted to prevent information leakage, while hyperparameters are selected using the validation set through grid search and early stopping when applicable. Forecasting performance is evaluated using multiple error metrics. Its forecasting performance is compared with several established forecasting approaches, including Seasonal Autoregressive Integrated Moving Average with Exogenous Regressors (SARIMAX), Poisson regression (PR), Negative Binomial regression (NBR), Extreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), and Long Short-Term Memory (LSTM). The results show that SARIMAX achieves the best overall performance in the univariate setting. In the multivariate setting, NBR and PR provide best performance. These findings indicate that the large-scale pretraining of TimeGPT do not necessarily translate into superior forecasting accuracy for emergency call volumes. Frugal methods trained solely on the available data remain better for this application.
Figures & tables
Figure 1: TimeGPT framework for emergency call forecasting, including pretraining, fine-tuning, and testing stages.
Method
Approximate number of trainable parameters
PR
d+1
NBR
d+2
SARIMAX
p+q+P+Q+d+1
XGBoost
O(∑m=1MNm)
MLP
(d+1)h1+∑k=1n−1(hkhk+1)+hno+∑k=1nhk+o
LSTM
4u(d+u+2)
Table 1: Conceptual comparison of the number of model parameters.
Model
Hyperparameter
Candidate values
TimeGPT
Fine-tuning steps
0,2,5,10,25,100,500
Fine-tuning loss
default, MAE, MAPE
Fine-tuning depth
1,2,3,4
LSTM
Number of LSTM units
32, 64, 128
Dense units
16, 32
Dropout
0.0, 0.2, 0.3
Table 2: Hyperparameter search space considered for the forecasting models.
Model
MAE (95% CI)
MAPE (%) (95% CI)
MASE
RMSE
RSD
Min AE
Max AE
P>10%
P>20%
TimeGPT
25.380 (23.514, 27.309)
8.462 (7.869, 9.067)
1.000
31.916
31.868
0.146
90.991
32.479
7.407
SARIMAX
19.281 (17.570, 21.157)
6.251 (5.771, 6.771)
0.760
24.831
24.353
0.018
83.917
18.519
1.425
NBR
25.803 (23.882, 27.868)
8.325 (7.771, 8.902)
1.017
32.372
31.102
0.404
109.563
34.188
3.989
PR
25.807 (23.884, 27.869)
8.326 (7.772, 8.903)
1.017
32.377
31.105
0.293
109.567
34.188
3.989
LSTM
27.338 (24.741, 30.341)
8.648 (7.981, 9.395)
1.078
35.181
32.075
0.007
117.110
34.758
6.268
MLP
25.632 (23.743, 27.657)
8.251 (7.731, 8.814)
1.010
32.418
31.112
0.181
103.951
31.909
5.413
Table 3: Performance comparison of the forecasting models for the univariate one-step-ahead scenario ( h=1 ).
Model
MAE (95% CI)
MAPE (%) (95% CI)
MASE
RMSE
RSD
Min AE
Max AE
P>10%
P>20%
TimeGPT
25.730 (23.619, 27.863)
8.562 (7.912, 9.243)
0.969
32.504
32.481
0.308
101.898
31.909
7.407
SARIMAX
22.602 (20.950, 24.314)
7.473 (6.960, 7.994)
0.851
28.888
28.926
0.059
94.883
26.211
4.843
NBR
27.811 (25.272, 30.518)
8.827 (8.159, 9.546)
1.047
35.354
32.120
0.323
115.156
39.316
6.838
PR
27.813 (25.273, 30.520)
8.827 (8.160, 9.546)
1.047
35.356
32.121
0.293
115.143
39.316
6.838
LSTM
26.943 (25.013, 28.986)
9.127 (8.421, 9.842)
1.014
33.664
32.493
0.008
100.090
35.328
8.547
MLP
27.733 (25.214, 30.383)
8.809 (8.172, 9.499)
1.044
35.256
32.368
0.128
109.297
35.897
6.268
Table 4: Performance comparison of the forecasting models for the univariate multi-step-ahead scenario ( h=7 ).
Figure 2: Univariate forecast MAPE evolution within the seven-day horizon.
Model
MAE (95% CI)
MAPE (%) (95% CI)
MASE
RMSE
RSD
Min AE
Max AE
P>10%
P>20%
TimeGPT
24.726 (22.863, 26.669)
8.111 (7.524, 8.685)
0.975
31.037
30.965
0.044
94.805
33.618
4.843
SARIMAX
20.307 (18.684, 22.062)
6.726 (6.223, 7.270)
0.800
25.508
25.535
0.114
81.018
24.217
1.994
NBR
23.146 (21.191, 25.218)
7.513 (6.973, 8.105)
0.912
30.121
29.848
0.318
119.157
28.205
3.989
PR
23.159 (21.214, 25.222)
7.521 (6.979, 8.106)
0.913
30.122
29.865
0.239
118.446
28.490
3.989
LSTM
25.163 (23.086, 27.407)
8.289 (7.637, 8.927)
0.992
32.311
32.336
0.046
104.507
30.199
6.553
MLP
37.078 (33.701, 40.688)
12.850 (11.607, 14.238)
1.462
44.212
33.600
0.299
119.175
53.276
18.519
Table 5: Performance comparison of the forecasting models for the multivariate one-step-ahead scenario ( h=1 ). Values in parentheses indicate 95% confidence intervals.
Model
MAE (95% CI)
MAPE (%) (95% CI)
MASE
RMSE
RSD
Min AE
Max AE
P>10%
P>20%
TimeGPT
24.930 (23.064, 26.856)
8.139 (7.618, 8.689)
0.938
31.465
31.297
0.242
107.117
34.188
5.698
SARIMAX
24.746 (22.871, 26.890)
8.202 (7.581, 8.888)
0.932
31.309
31.338
0.197
104.921
32.479
5.128
NBR
23.439 (21.180, 25.898)
7.568 (6.956, 8.223)
0.882
31.037
30.521
0.181
129.741
28.490
4.843
PR
23.425 (21.184, 25.874)
7.567 (6.957, 8.223)
0.882
31.012
30.524
0.268
128.978
28.490
5.128
LSTM
43.694 (39.530, 48.106)
13.559 (12.450, 14.662)
1.645
52.646
32.252
0.211
165.086
66.667
19.373
MLP
607.185 (599.292, 614.930)
201.437 (197.011, 206.048)
22.857
630.168
168.875
292.804
973.597
100.000
100.000
Table 6: Performance comparison of the forecasting models for the multivariate multi-step-ahead scenario ( h=7 ). Values in parentheses indicate 95% confidence intervals.
Figure 3: Multivariate forecast MAPE evolution within the seven-day horizon.
Forecast
Horizon
TimeGPT vs SARIMAX
TimeGPT vs MA
SARIMAX vs MA
Univariate
h=1
<0.0001
0.9817
<0.0001
h=7
<0.0001
0.2588
<0.0001
Multivariate
h=1
<0.0001
0.3881
<0.0001
h=7
0.8460
0.0535
0.0424
Table 7: Pairwise Diebold–Mariano test p-values for univariate and multivariate forecasting settings. Bold values indicate statistically significant differences ( p<0.05 ).
Time series forecasting predicts future values from ordered observations using statistical, neural, and probabilistic models to support decisions under uncertainty.
This study investigates whether generalized means can achieve forecasting performance comparable to that of TimeGPT for daily emergency call volume prediction. Specifically, two families of generalized means, the Hölder means (HM) and the Lehmer means (LM), are evaluated against fine-tuned TimeGPT across forecasting horizons ranging from one to seven days. Forecasting performance is assessed on a real-world emergency call dataset using multiple error metrics, and Diebold-Mariano test. Experimental results show that HM and LM achieve nearly identical predictive performance across all forecasting horizons. For short-term forecasting, both generalized means outperform TimeGPT, with reductions of mean absolute error (MAE) approximately 1.24% and 2.48% at horizons of one and three days, respectively. Conversely, TimeGPT achieves a lower error at the seven-day horizon, with an descresing of MAE by 1.98%. However, Diebold-Mariano tests reveal no statistically significant differences between the methods at the evaluated horizons. These findings suggest that generalized means can provide competitive forecasting performance compared with a energyvore foundation model.
Accurate forecasting with uncertainty quantification is critical in a variety of scientific applications ranging from epidemiology and economics to energy systems and earth sciences. Yet most forecasting models still need to be developed and trained separately for each domain-specific dataset. Time series foundation models (TSFMs) aim to remove this limitation and enable the application of the same model across domains through large-scale pretraining. However, most existing TSFMs treat each time series independently and therefore cannot model interactions among time series and external drivers. Here, we introduce Chronos-2, a foundation model for zero-shot forecasting that models the interactions between multivariate time series and heterogeneous covariates, which commonly arise in scientific applications. The model infers relationships among time series directly from historical context and generates probabilistic forecasts without task-specific treatment. Across various benchmarks, Chronos-2 achieves state-of-the-art accuracy on average by improving skill scores by 5.2 percentage points relative to the strongest baselines and increasing win rates by 10 percentage points. Chronos-2 is also computationally efficient, generating forecasts for 100 time series in under one second on a single mid-range GPU, 50x and 1000x faster end-to-end than the best task-specific point and probabilistic forecasters, respectively. In addition, we demonstrate its effectiveness in diverse scientific settings -- including infectious-disease forecasting, macroeconomic modeling, carbon-flux prediction, and runoff simulation -- where the model matches or surpasses specialized methods developed for each domain. These results show that a single pretrained model can capture structure shared across temporal systems.
Abdul Fatir Ansari, Oleksandr Shchur, Jaris Küken +22
Amazon Web Services · University of Freiburg · Amazon +3
We introduce a new context-enriched, multimodal time series forecasting benchmark, TimesX. TimesX contains a wide selection of high-quality real-world time series with diverse domains and textual contexts obtained from an automated data generation pipeline, which helps address three main issues of existing multimodal forecasting benchmarks: (1) poor generalization due to the small scale and synthetic nature of benchmark data, (2) very limited types of textual contexts in the benchmarks, and (3) an inability to mitigate data leakage in evaluation. We conduct a thorough empirical study of zero-shot multimodal forecasting approaches on TimesX. Our results suggest that many approaches that perform well on existing benchmarks may fail on TimesX. In contrast, simple ensemble methods that leverage rich textual context accompanying time-series can outperform strong baselines on TimesX.