Time Series Foundation Models (TSFMs) currently provide state-of-the-art results in forecasting tasks. They are available out-of-the-box and rely on in-context learning to make their predictions, which makes the quality of their performance highly sensitive to the user-selected lookback, covariates, horizon and training data distributions. In practise, the quality of the forecasts are variable but complementary, which highlights the need for a principled ensembling approach, rather than selecting the best context. This paper introduces Latent Inference-Time Guidance for TSFMs, which adaptively combines a pool of TSFM forecasts through a time-dependent latent space with independent components. The framework comes equipped with identifiability and reconstruction guarantees, whilst maintaining the off-the-shelf aspect of foundation models. We provide experiments on datasets at various frequencies and from multiple domains: these show that the approach is competitive with traditional ensembling approaches.
Figures & tables
Dataset
Domain
Freq
Length
# Covariates
# Series
London Smart Meter ( UK Power Networks, 2015 )
Energy
D
606
17
5
Mauna Loa ( Lan and Keeling, 2026 )
Climate
M
627
9
1
NN5 Weekly ( Godahewa et al., 2020 )
Finance
W
837
4
111
Washington Bicycle Share ( Fanaee-T, 2013 )
Traffic
H
17378
30
1
Table 1: Datasets used to evaluate Lit i g - Tsfm and their key characteristics.
Figure 1: Average forecasting performance ×102 on datasets with strong TSFM forecasts. Lower is better. Note that the y axes are on different scales.
Figure 2: Average forecasting performance ×102 on datasets with poor TSFM forecasts: for the Washington Bicycle Share dataset, mean RMSE (std) is 19.45 (6.61) and mean MAE (std) is 14.70 (5.65); for the Mauna Loa dataset, mean RMSE (std) is 45.27 (17.33) and mean MAE (std) is 35.34 (13.84). Lower is better. Note that the y axes are on different scales.
Figure 3: Experts ablation: Lit i g - Tsfm performance compared to baseline ensemble methods on the Washington Bicycle Share dataset, with nine strong experts and one, three and five poor ones respectively. Metrics given ×10−2 , lower is better.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Feature
Preprocessing
Acorn segment
One-hot encoded
Number of clients
Normalised between 0 and 1
Date
Year cyclically encoded
Weekday
One-hot encoded
Temperature
Normalised between 0 and 1
Consumption (target)
Normalised between 0 and 1
Appendix
Table 2: Preprocessing for variables in the London Smart meter dataset.
Feature
Preprocessing
Demand (target)
Normalised between 0 and 1
Month of year
Cyclically encoded
Week of year
Cyclically encoded
Appendix
Table 3: Preprocessing for variables in the NN5 weekly dataset.
Feature
Preprocessing
Positional encoding
Encoded between 0 and 1
Year
Normalised between 0 and 1
Month
Cyclically encoded
Mean monthly concentration (target)
Detrended, normalised between 0 and 1; lagged
Deseasonalised mean monthly concentration
Detrended, normalised between 0 and 1; lagged
Number of days with measurements
Normalised between 0 and 1
Appendix
Table 4: Preprocessing for variables in the Mauna Loa monthly dataset.
Feature
Preprocessing
Season
One-hot encoded
Bank holiday
One-hot encoded
Day of week
One-hot encoded
Hour
Cyclically encoded
Month
Cyclically encoded
Working day
One-hot encoded
Appendix
Table 5: Preprocessing for variables in the Washington Bicycle Share dataset.
Hyperparameter
Smart Meter
Bike Share
NN5 Weekly
Mauna Loa
Sequence length
64
96
16
196
Batch size
128
256
64
128
Learning rate
5×10−3
1×10−3
1×10−3
5×10−3
Latent dimension
64
64
32
64
Epochs
200
20
60
500
Number heads in encoder
4
4
4
4
Appendix
Table 6: Hyperparameters chosen for each dataset.
Dataset
TSFM low freq.
TSFM high freq.
Kalman
Lit i g - Tsfm train + infer
Smart Meter*
1’44, 1’27
Not used
35”
3’04” + 1”
Bike Share
39’, 10’
60’, 10’
13’02"
9’48" + 33"
NN5*
10", 9"
40", 59"
5"
37" + 1"
Mauna Loa
1’, 1’
14’, 16’
1’
4’ + 2”
Appendix
Table 7: Compute cost for Lit i g - Tsfm compared to the baseline models. Datasets with a * show cost for fitting the Kalman Filter and running Lit i g - Tsfm inference a single series. TSFM prompting times show the time needed to make a mean forecast and a raw quantiles forecast.
Figure 4: London Smart Meter, Acorn Segment K forecast sample of Lit i g - Tsfm and its baselines.
Figure 5: Washington Bike Share forecast sample of Lit i g - Tsfm and its baselines.
Figure 6: NN5 forecast sample of Lit i g - Tsfm and its baselines.
Figure 7: Detrended Mauna Loa forecast sample of Lit i g - Tsfm and its baselines.
Figure 8: Architecture ablation: Lit i g - Tsfm performance on the London Smart Meter dataset with different encoder and decoders. Metrics given ×10−2 , lower is better.