STRATA-TS: Target-Aware Urban Forecasting

• STRATA-TS is a cross-city time series forecasting framework that selectively transfers source subsequences using target-aware retrieval.
• It employs a patch-based temporal encoder and LLM-guided reasoning to combine source data with limited target observations for improved predictions.
• Experimental results demonstrate significant improvements in MAE and MAPE, validating its effectiveness in urban forecasting with sparse data.

STRATA-TS, short for Selective TRAnsfer via TArget-aware retrieval for Time Series, is a cross-city time series forecasting framework for urban forecasting in data-scarce regimes. It addresses the setting in which a data-rich source city and data-scarce target cities exhibit severe imbalance in observation density and history length, and it does so by combining target-aware retrieval of source subsequences with LLM-based reasoning and distillation. In the formulation described for parking availability forecasting, STRATA-TS uses Singapore as the source city and Nottingham and Glasgow as target cities, operates on 15-minute interval data, retrieves source patterns that are semantically and dynamically aligned with a target query, and uses those retrieved exemplars to guide a reasoning stage whose behavior is later distilled into a compact open model (Jiang et al., 26 Aug 2025).

1. Problem formulation and motivation

The framework is motivated by urban time series forecasting, instantiated as parking availability forecasting across multiple cities. Each city contains multiple parking locations monitored at 15-minute intervals. If $N$ denotes the number of nodes and $L$ the length of a time sequence segment, a multivariate batch is written as

$\mathbf{X}\in\mathbb{R}^{B\times N\times L},$

where $B$ is batch size, $N$ is the number of nodes, and $L$ is the number of time steps. Following GTS, the framework uses

$L_{\text{in}} = L_{\text{out}} = 12,$

so the forecasting task is to use 3 hours of history to predict the next 3 hours (Jiang et al., 26 Aug 2025).

For a city $c$, node $n$, and time $t$, the objective is: $L$0 The source city, Singapore, is described as data-rich, with $L$1 carparks and 813,984 records over a period from 1 May 2021 – 6 June 2021. Nottingham and Glasgow are data-scarce targets with 11 carparks and 7 carparks, respectively, over 26 Oct 2016 – 16 Feb 2017. In the transfer setting, only three days of training data are used for target cities (Jiang et al., 26 Aug 2025).

The central motivation is that naive cross-city transfer is unreliable. Only a limited subset of source-city subsequences is actually useful for a target query, while indiscriminate transfer introduces noise and can cause negative transfer. The data block reports that when similarity weights $L$2 were computed between a Nottingham query slice and randomly sampled Singapore slices, only <5% of source slices have $L$3, implying that roughly 95% of potential source information is noisy for that target query (Jiang et al., 26 Aug 2025).

Dataset	Role	Key statistics
Singapore	Source city	139 carparks, 813,984 records, 15 mins
Nottingham	Target city	11 carparks, 15 mins, 3-day target training
Glasgow	Target city	7 carparks, 15 mins, 3-day target training

2. System architecture

STRATA-TS consists of two main modules: a Target-Aware Retriever Module and a Retrieval-Guided Reasoning Module. The retriever learns segment-level embeddings from source-city time series, retrieves top-$L$4 aligned exemplars for a target query, and supplies both numerical sequences and textual metadata to the reasoning stage. The reasoning module combines a base spatio-temporal predictor with a reasoning-capable large model, then distills the resulting behavior into a compact student model suitable for deployment (Jiang et al., 26 Aug 2025).

The retriever side starts by pretraining a patch-based temporal encoder on source-city data using a Masked Autoencoder. A source-domain knowledge base is then built, storing for each segment an identifier, an embedding $L$5, a long-term source sequence $L$6, and associated text context $L$7: $L$8 The knowledge base is reported to contain over 500k segments (Jiang et al., 26 Aug 2025).

The reasoning side uses TransGTR as the base spatio-temporal model, trained on source plus limited target data, to generate numerical predictions. A powerful teacher LLM, specifically GPT-o1 / DeepSeek-R1, is then used during training to perform structured reasoning over target history, retrieved source sequences, source and target textual context, and base-model predictions. The reasoning patterns are distilled into LLaMA-3-8B through supervised fine-tuning, with 4-bit quantization and LoRA of rank $L$9, using Adam with learning rate $\mathbf{X}\in\mathbb{R}^{B\times N\times L},$0 for 5 epochs (Jiang et al., 26 Aug 2025).

At inference time, the system does not depend on the proprietary teacher. Instead, it uses the target-aware retriever, the base predictor, and the distilled LLaMA-3-8B model to produce final forecasts.

3. Target-aware retrieval mechanism

The retrieval module is built around a patch-based temporal encoder. Given

$\mathbf{X}\in\mathbb{R}^{B\times N\times L},$1

the input is partitioned into non-overlapping patches of width $\mathbf{X}\in\mathbb{R}^{B\times N\times L},$2, yielding $\mathbf{X}\in\mathbb{R}^{B\times N\times L},$3 patches: $\mathbf{X}\in\mathbb{R}^{B\times N\times L},$4 Each patch is projected into a $\mathbf{X}\in\mathbb{R}^{B\times N\times L},$5-dimensional embedding, and instance normalization is applied: $\mathbf{X}\in\mathbb{R}^{B\times N\times L},$6 The encoder then reshapes the representation to

$\mathbf{X}\in\mathbb{R}^{B\times N\times L},$7

applies a spatial CNN,

$\mathbf{X}\in\mathbb{R}^{B\times N\times L},$8

and feeds the result to a stack of Transformer decoder layers in a Masked Autoencoder setup. The MAE reconstruction loss is

$\mathbf{X}\in\mathbb{R}^{B\times N\times L},$9

After pretraining, the decoder is discarded and the encoder is used as the segment embedding function (Jiang et al., 26 Aug 2025).

For a target query sequence, the encoder produces a query embedding $B$0. Both $B$1 and a source embedding $B$2 are flattened to vectors $B$3 and $B$4, and similarity is computed using both Euclidean and covariance-aware geometry: $B$5

$B$6

$B$7

A normalized similarity weight used for visualization is

$B$8

Top-$B$9 retrieval is then defined as

$N$0

This retrieval mechanism is the basis of the framework’s selective transfer claim: it transfers only those source subsequences whose embeddings are strongly aligned with the target query, rather than transferring source-city structure wholesale. The retrieved item includes both the long-term sequence $N$1 and text context $N$2, allowing downstream reasoning to use both temporal and semantic alignment (Jiang et al., 26 Aug 2025).

4. Retrieval-guided reasoning and distillation

The reasoning module takes the retrieved exemplars and combines them with a base spatio-temporal model. In implementation, TransGTR produces predictions for all target nodes: $N$3 with $N$4 (Jiang et al., 26 Aug 2025).

During training, GPT-o1 or DeepSeek-R1 is prompted with a structured input containing: source textual information, source long-term sequence, prediction horizon, target textual information, target historical records, base-model predictions, and ground truth. The instruction is not to output future parking values, but instead to extract useful hints relating source context, target history, simulation predictions, and ground truth. The reasoning stage is described as deriving causal and corrective hints, including bias corrections, lag and seasonality cues, and cross-city alignment (Jiang et al., 26 Aug 2025).

A representative prompt structure includes fields such as “Source city textual information,” “Source long-term sequence,” “Target city textual information,” “Target historical records,” “Simulation predictions,” and “ground truth,” together with explicit rules stating that the model should use the ground truth only to infer relationships rather than to reproduce it. The teacher model’s output is then used as a supervised target for LLaMA-3-8B, which learns to emulate retrieval-guided reasoning in a compact open model (Jiang et al., 26 Aug 2025).

At inference time, ground truth is absent. The base ST model produces $N$5, the retriever returns $N$6, and the distilled student model performs structured reasoning to output the final forecast. The overall forecasting function is summarized as

$N$7

This makes the system modular: numerical forecasting is provided by a dedicated spatio-temporal backbone, while cross-city adjustment and interpretation are handled by retrieval-guided reasoning (Jiang et al., 26 Aug 2025).

5. Experimental results and ablations

The framework is evaluated against general time-series baselines, spatio-temporal graph models, cross-city transfer models, and LLM-based time-series models. The reported baselines include DLinear, PatchTST, GPT4TS, Time-LLM, Graph WaveNet (GWN), GTS, ST-GFSL, TPB, and TransGTR (Jiang et al., 26 Aug 2025).

On Nottingham, STRATA-TS achieves the best performance in 32 out of 36 metric/horizon combinations. At the 15 min horizon, it reports MAE = 7.09 and MAPE = 4.73%, compared with TransGTR: MAE = 8.14, MAPE = 6.56%. At 120 min, it reports MAE = 30.94 and MAPE = 23.63%, compared with TransGTR: MAE = 33.35, MAPE = 30.09%. At 180 min, it reports MAE = 44.08 and MAPE = 35.16%, compared with TransGTR: MAE = 46.61, MAPE = 44.73% (Jiang et al., 26 Aug 2025).

On Glasgow, STRATA-TS achieves the best performance in 34 out of 36 metric/horizon combinations. At 15 min, it reports MAE = 7.50 and MAPE = 1.50%, compared with TransGTR: MAE = 9.84, MAPE = 1.96%. At 120 min, it reports MAE = 34.25 and MAPE = 6.81%, compared with TransGTR: MAE = 36.14, MAPE = 7.09%. At 180 min, it reports MAE = 45.60 and MAPE = 9.15%, compared with TransGTR: MAE = 50.92, MAPE = 10.10%. Across both target datasets, the framework wins 66 of 72 evaluations (Jiang et al., 26 Aug 2025).

The ablation study isolates two components: Random-Centroid Retrieval, which replaces similarity-based retrieval with k-means clustering and random centroid selection, and Open-Source LLM Inference Only, which replaces teacher-guided reasoning with a directly deployed LLaMA-3-8B. Both variants show noticeable drops in MAE and MAPE, supporting two specific conclusions: retrieval precision is critical, and teacher-guided reasoning followed by distillation contributes materially beyond the direct use of the open model (Jiang et al., 26 Aug 2025).

The paper also reports qualitative evidence. Retrieval case studies show that after z-normalization, the retrieved source sequence aligns closely with the target query, while nearby source sequences deviate. Forecast visualizations show instances where STRATA-TS tracks declining availability or preserves stable availability more accurately than TransGTR. These analyses support the claim that retrieval is not merely nearest-neighbor reuse, but a selective transfer mechanism coupled to reasoning (Jiang et al., 26 Aug 2025).

6. Positioning, interpretation, and limitations

Within the cross-city forecasting literature, STRATA-TS is positioned against transfer-learning and meta-learning methods such as RegionTrans, CrossTReS, MetaST, ST-GFSL, DASTNet, and TransGTR. The distinction claimed for STRATA-TS is that it operates at the subsequence level rather than transferring whole source-city models, graphs, or global structures, and that it combines retrieval with LLM reasoning rather than relying solely on parametric adaptation (Jiang et al., 26 Aug 2025).

Within LLM-based time-series forecasting, it is contrasted with GPT4TS, LLM4TS, and Time-LLM, which reprogram numerical time series into tokens and typically treat the problem as direct prediction. STRATA-TS instead uses an LLM as a reasoning layer, while keeping numerical prediction in a dedicated spatio-temporal backbone. It is also distinguished from retrieval-augmented forecasting approaches such as TimeRAG, RAF, TS-RAG, RATD, and RAFT, because its retrieval uses a domain-adapted MAE-pretrained patch encoder and because retrieval is coupled to a teacher-student reasoning pipeline rather than simple concatenation of retrieved sequences (Jiang et al., 26 Aug 2025).

The framework’s interpretability claims arise from three levels of visibility: retrieval heatmaps that show which source slices are selected, case studies in which retrieved sequences visibly match target patterns, and reasoning traces that describe how source and target dynamics, textual semantics, and base predictions interact. This suggests an interpretable path for transfer: retrieved exemplars, reasoning over those exemplars, then a forecast corrected by that reasoning (Jiang et al., 26 Aug 2025).

The reported limitations are also specific. The approach depends on the quality of the teacher LLM; pretraining the MAE encoder and maintaining a knowledge base of over 500k segments is computationally nontrivial; and the method assumes that some source-city segments are semantically and dynamically similar to target-city queries. The paper also notes an implicit stationarity assumption in the learned retrieval embeddings, so strong non-stationarities may degrade retrieval quality. Proposed future directions include extending the framework to multi-modal urban data, more scalable retrieval systems, and other spatio-temporal tasks beyond parking, such as traffic speeds, energy demand, and mobility flows (Jiang et al., 26 Aug 2025).