AutoHFormer: Efficient Time-Series Transformer

• The paper presents a hierarchical autoregressive modeling framework that decomposes forecasting into segment-level blocks to refine predictions and reduce error accumulation.
• AutoHFormer introduces dynamic windowed masked attention with an exponential decay mechanism to enforce strict temporal causality and improve computational efficiency.
• Empirical evaluations demonstrate that AutoHFormer achieves significant speedups, lower memory usage, and improved accuracy compared to full attention models on benchmark datasets.

AutoHFormer is a Transformer-based architecture for long-horizon time-series forecasting that simultaneously enforces strict temporal causality, achieves sub-quadratic complexity, and captures multi-scale temporal patterns. It is characterized by its hierarchical autoregressive generative mechanism, dynamic windowed masked attention with exponential decay, and a hybrid temporal encoding scheme. These design elements collectively address the fundamental requirements for reliable, scalable, and precise time-series prediction (Zhang et al., 19 Jun 2025).

1. Hierarchical Autoregressive Modeling Framework

AutoHFormer adopts a hierarchical generation scheme that decomposes the forecasting problem into segment-level blocks, with each block processed via an initial summary followed by step-wise autoregressive refinement. Given an input history $X_{1:L}\in\mathbb{R}^{L\times V}$ and total forecast length $T_{\text{total}}=K\cdot H$ (where $K$ is the number of segments and $H$ is the segment length), the model factorizes the predictive distribution as

$$p(\hat{Y} \mid X_{1:L}) = \prod_{h=1}^K p(\hat{Y}_h \mid X_{1:L},\hat{Y}_{1:h-1}) = \prod_{h=1}^K \prod_{t=1}^H p(\hat{y}_{(h-1)H+t} \mid X_{1:L},\hat{Y}_{1:h-1},\hat{y}_{(h-1)H+1:(h-1)H+t-1})$$

Segment-level initialization is executed through a block-wise Transformer: $$\hat{Y}_h^{\text{init}} = \mathcal{F}_\theta(C_h) \in \mathbb{R}^{H \times d}$$ where $$C_h = \text{Concat}(X_{1:L}, \hat{Y}_{1:h-1}) \in \mathbb{R}^{(L + (h-1)H) \times d}$$. Within each segment $h$, step-wise refinement generates each scalar forecast recursively: $$C_h^t = \text{Concat}(C_h, \hat{y}_{(h-1)H+1}, \ldots, \hat{y}_{(h-1)H+t-1})$$

$$o_t = \text{WindowedAttention}(C_h^t, C_h^t, C_h^t;A_t)$$

$T_{\text{total}}=K\cdot H$0

$T_{\text{total}}=K\cdot H$1

This hybrid approach reduces error accumulation typical of sequential autoregression by correcting each step with windowed context and parameter sharing across segment steps.

2. Dynamic Windowed Masked Attention Mechanism

AutoHFormer replaces full self-attention with Dynamic Windowed Masked Attention (DWMA), wherein at each time step $T_{\text{total}}=K\cdot H$2 (of $T_{\text{total}}=K\cdot H$3), attention is restricted to a causal window: $T_{\text{total}}=K\cdot H$4 No anti-causal (future) information is leaked. Within this window, attention between $T_{\text{total}}=K\cdot H$5 and $T_{\text{total}}=K\cdot H$6 is modulated by an exponential decay kernel with a learnable rate $T_{\text{total}}=K\cdot H$7: $T_{\text{total}}=K\cdot H$8 The attention score matrix thus becomes: $T_{\text{total}}=K\cdot H$9 where $K$0 and $K$1 are the relative-position encodings. As $K$2 positions attend to at most $K$3 others, the time and space complexity is reduced to $K$4, significantly below the $K$5 of vanilla self-attention.

3. Adaptive Temporal Encoding

The model incorporates a hybrid positional encoding to simultaneously model short-term transients and long-term temporal dynamics. For position pair $K$6, a fixed sinusoidal term is precomputed: $K$7 for $K$8. These are stored in a lookup table and provide continuous, shift-invariant encodings across lags. The learnable decay parameter $K$9 is trained jointly, allowing the effective receptive field to adapt per dataset by tuning the influence of distal context: $H$0

4. Enforcement of Strict Temporal Causality

Temporal causality is strictly enforced by two complementary mechanisms:

• A hard attention mask $H$1 zeroes all matrix entries where $H$2 exceeds $H$3 in both the segment-level and intra-segment passes, eliminating anti-causal dependencies.
• The causal window $H$4 only includes current and preceding time indices by construction.

Consequently, every forecast $H$5 is a measurable function of only $H$6 and $H$7, preserving $H$8.

5. Computational and Memory Efficiency

AutoHFormer achieves sub-quadratic runtime and space scaling owing to its windowed attention. The key complexities for various models are summarized as follows:

Model	Time Complexity	Space Complexity
Full Attention	$H$9	$$p(\hat{Y} \mid X_{1:L}) = \prod_{h=1}^K p(\hat{Y}_h \mid X_{1:L},\hat{Y}_{1:h-1}) = \prod_{h=1}^K \prod_{t=1}^H p(\hat{y}_{(h-1)H+t} \mid X_{1:L},\hat{Y}_{1:h-1},\hat{y}_{(h-1)H+1:(h-1)H+t-1})$$0
AutoHFormer	$$p(\hat{Y} \mid X_{1:L}) = \prod_{h=1}^K p(\hat{Y}_h \mid X_{1:L},\hat{Y}_{1:h-1}) = \prod_{h=1}^K \prod_{t=1}^H p(\hat{y}_{(h-1)H+t} \mid X_{1:L},\hat{Y}_{1:h-1},\hat{y}_{(h-1)H+1:(h-1)H+t-1})$$1	$$p(\hat{Y} \mid X_{1:L}) = \prod_{h=1}^K p(\hat{Y}_h \mid X_{1:L},\hat{Y}_{1:h-1}) = \prod_{h=1}^K \prod_{t=1}^H p(\hat{y}_{(h-1)H+t} \mid X_{1:L},\hat{Y}_{1:h-1},\hat{y}_{(h-1)H+1:(h-1)H+t-1})$$2
RNN-based	$$p(\hat{Y} \mid X_{1:L}) = \prod_{h=1}^K p(\hat{Y}_h \mid X_{1:L},\hat{Y}_{1:h-1}) = \prod_{h=1}^K \prod_{t=1}^H p(\hat{y}_{(h-1)H+t} \mid X_{1:L},\hat{Y}_{1:h-1},\hat{y}_{(h-1)H+1:(h-1)H+t-1})$$3	$$p(\hat{Y} \mid X_{1:L}) = \prod_{h=1}^K p(\hat{Y}_h \mid X_{1:L},\hat{Y}_{1:h-1}) = \prod_{h=1}^K \prod_{t=1}^H p(\hat{y}_{(h-1)H+t} \mid X_{1:L},\hat{Y}_{1:h-1},\hat{y}_{(h-1)H+1:(h-1)H+t-1})$$4

With $$p(\hat{Y} \mid X_{1:L}) = \prod_{h=1}^K p(\hat{Y}_h \mid X_{1:L},\hat{Y}_{1:h-1}) = \prod_{h=1}^K \prod_{t=1}^H p(\hat{y}_{(h-1)H+t} \mid X_{1:L},\hat{Y}_{1:h-1},\hat{y}_{(h-1)H+1:(h-1)H+t-1})$$5, $$p(\hat{Y} \mid X_{1:L}) = \prod_{h=1}^K p(\hat{Y}_h \mid X_{1:L},\hat{Y}_{1:h-1}) = \prod_{h=1}^K \prod_{t=1}^H p(\hat{y}_{(h-1)H+t} \mid X_{1:L},\hat{Y}_{1:h-1},\hat{y}_{(h-1)H+1:(h-1)H+t-1})$$6, $$p(\hat{Y} \mid X_{1:L}) = \prod_{h=1}^K p(\hat{Y}_h \mid X_{1:L},\hat{Y}_{1:h-1}) = \prod_{h=1}^K \prod_{t=1}^H p(\hat{y}_{(h-1)H+t} \mid X_{1:L},\hat{Y}_{1:h-1},\hat{y}_{(h-1)H+1:(h-1)H+t-1})$$7, AutoHFormer achieves up to $$p(\hat{Y} \mid X_{1:L}) = \prod_{h=1}^K p(\hat{Y}_h \mid X_{1:L},\hat{Y}_{1:h-1}) = \prod_{h=1}^K \prod_{t=1}^H p(\hat{y}_{(h-1)H+t} \mid X_{1:L},\hat{Y}_{1:h-1},\hat{y}_{(h-1)H+1:(h-1)H+t-1})$$8 speedups over full self-attention. On the PEMS08 dataset with $$p(\hat{Y} \mid X_{1:L}) = \prod_{h=1}^K p(\hat{Y}_h \mid X_{1:L},\hat{Y}_{1:h-1}) = \prod_{h=1}^K \prod_{t=1}^H p(\hat{y}_{(h-1)H+t} \mid X_{1:L},\hat{Y}_{1:h-1},\hat{y}_{(h-1)H+1:(h-1)H+t-1})$$9, it is $$\hat{Y}_h^{\text{init}} = \mathcal{F}_\theta(C_h) \in \mathbb{R}^{H \times d}$$0 faster per epoch (4.58s vs. 49.3s) and uses $$\hat{Y}_h^{\text{init}} = \mathcal{F}_\theta(C_h) \in \mathbb{R}^{H \times d}$$1 less GPU memory (2.99 GB vs. 18.13 GB) than PatchTST, while maintaining or improving predictive accuracy (Zhang et al., 19 Jun 2025).

6. Empirical Performance and Benchmark Results

Comprehensive evaluations were performed on eight public benchmarks: ETTh1/2, ETTm1/2, PEMS04/08, Weather, and Electricity datasets, and four horizons $$\hat{Y}_h^{\text{init}} = \mathcal{F}_\theta(C_h) \in \mathbb{R}^{H \times d}$$2. Under autoregressive evaluation, AutoHFormer achieved first rank in $$\hat{Y}_h^{\text{init}} = \mathcal{F}_\theta(C_h) \in \mathbb{R}^{H \times d}$$3 cases. On PEMS08, it reported MSE/MAE of $$\hat{Y}_h^{\text{init}} = \mathcal{F}_\theta(C_h) \in \mathbb{R}^{H \times d}$$4 (versus PatchTST’s $$\hat{Y}_h^{\text{init}} = \mathcal{F}_\theta(C_h) \in \mathbb{R}^{H \times d}$$5), indicating $$\hat{Y}_h^{\text{init}} = \mathcal{F}_\theta(C_h) \in \mathbb{R}^{H \times d}$$6 and $$\hat{Y}_h^{\text{init}} = \mathcal{F}_\theta(C_h) \in \mathbb{R}^{H \times d}$$7 relative improvement. Hierarchical loss is defined by

$$\hat{Y}_h^{\text{init}} = \mathcal{F}_\theta(C_h) \in \mathbb{R}^{H \times d}$$8

with segment discount $$\hat{Y}_h^{\text{init}} = \mathcal{F}_\theta(C_h) \in \mathbb{R}^{H \times d}$$9, improving long-term prediction stability.

7. Limitations and Open Issues

Several limitations and avenues for future work exist:

• Fixed Window Bound: Setting window size $$C_h = \text{Concat}(X_{1:L}, \hat{Y}_{1:h-1}) \in \mathbb{R}^{(L + (h-1)H) \times d}$$0 trades off context coverage against efficiency. Small $$C_h = \text{Concat}(X_{1:L}, \hat{Y}_{1:h-1}) \in \mathbb{R}^{(L + (h-1)H) \times d}$$1 may omit long-range dependencies.
• Decay Parameter Sensitivity: Careful tuning of $$C_h = \text{Concat}(X_{1:L}, \hat{Y}_{1:h-1}) \in \mathbb{R}^{(L + (h-1)H) \times d}$$2 is required. Large values flatten the decay, reducing discriminative context; small values focus attention on immediate past, possibly ignoring useful historical patterns.
• Streaming and Non-stationarity: The architecture assumes a fixed look-back $$C_h = \text{Concat}(X_{1:L}, \hat{Y}_{1:h-1}) \in \mathbb{R}^{(L + (h-1)H) \times d}$$3 and stationarity. Extending to streaming or non-stationary environments is an open research question.
• Hierarchical Depth and Interpretability: Additional hierarchical layers may enhance expressivity, but their role in multi-scale pattern discovery merits further theoretical analysis.

In summary, AutoHFormer introduces a principled framework for strictly causal, efficient, and accurate long-horizon time-series forecasting by integrating hierarchical autoregressive modeling, dynamic windowed attention with decay, and adaptive temporal encodings (Zhang et al., 19 Jun 2025).