Time-series Dense Encoder (TiDE) Overview

• TiDE is a deep neural network family that unifies encoder–decoder frameworks, sinusoidal time encoding, and dense MLP architectures for advanced time-series modeling.
• It applies to irregular sampling, long-horizon forecasting, robust control, and reinforcement learning, effectively integrating covariates and handling missingness.
• Empirical evaluations show TiDE’s efficiency in forecasting and control, offering lower prediction errors and faster compute times than traditional Transformer and RNN models.

The Time-series Dense Encoder (TiDE) encompasses a family of deep neural architectures for time-series modeling, unifying recent advances in encoder–decoder frameworks, sinusoidal time encoding, and dense multi-layer perceptron (MLP) designs. TiDE has been instantiated for both sequence modeling under irregular sampling and, more broadly, for long-horizon sequence forecasting, robust control, and reinforcement learning. Canonical features include MLP-based residual blocks, rapid multi-step sequence encoding, and direct handling of both covariates and missingness patterns, with proven empirical and theoretical advantages in domains ranging from clinical risk prediction to digital twins and financial decision processes.

1. Core Architectural Designs

Two primary lines can be traced in the literature for TiDE:

• Sinusoidal Time Embedding TiDE: The original instantiation for irregularly sampled series augments each observation with a deterministic, fixed-dimensional sinusoidal embedding of its timestamp. At each step $t$, a $d_{\mathrm{TE}}$-dimensional time embedding $TE(t)$ is computed with components:

$$\begin{aligned} TE(t)_{2i} &= \sin\left( t / \text{maxtime}^{2i/d_{\mathrm{TE}}} \right) \ TE(t)_{2i+1} &= \cos\left( t / \text{maxtime}^{2i/d_{\mathrm{TE}}} \right) \end{aligned}$$

Typically, $d_{\mathrm{TE}} = 32$, $\text{maxtime}=48$h (Sousa et al., 2020).

• MLP Encoder-Decoder TiDE: For fully data-driven forecasting, TiDE uses a hierarchical stack of residual MLP ("ResidualBlock") layers, flattening time-steps and covariates, merging static and dynamic features, and passing through encoder and decoder MLPs. Feature-projection blocks process covariates, while a "temporal decoder" fuses per-horizon latent representations with step-specific covariates. A global linear residual module ensures all linear autoregressive solutions are contained (Das et al., 2023).
• General ResNet MLP TiDE: For control and reinforcement learning, variants flatten multivariate history, propagate through residual dense blocks with layer normalization, then project to a compact encoding fed to downstream policy or value networks (Liu et al., 12 Aug 2025, Chen et al., 17 Jan 2025, Chen et al., 10 Jan 2025).

Common features across instances are illustrated below:

Variant	Input Handling	Positional/Time Encoding	Output Functionality
Sinusoidal TiDE	$x(t)\in\mathbb{R}^D$	Sinusoid (fixed, concat/add)	Per-step input to RNN/MLP
Encoder-Decoder TiDE	$y_{1:L},x_{1:L+H},s$	Order-provided (no explicit)	One-shot multi-horizon prediction
RL/Control TiDE	Flattened window, covariate	Flattened/implicit	State encoding for agent/control

2. Mathematical Formulation

Sinusoidal Time Embeddings (Irregular Sampling)

Given a series of irregular events at times $\{t_j\}$:

• Construct at each $t_j$ a time embedding $d_{\mathrm{TE}}$0 as above.
• Integrate via concatenation with features or addition to hidden state, e.g.,
  • $d_{\mathrm{TE}}$1 input to LSTM (catTE), or,
  • $d_{\mathrm{TE}}$2 after processing in self-attentive LSTM (addTE).
• Downstream MLPs pool encoded sequence for output (e.g., mortality risk, length of stay) (Sousa et al., 2020).

Encoder–Decoder MLP (Long-Term Forecasting)

Formally, for sample $d_{\mathrm{TE}}$3:

• Inputs: history $d_{\mathrm{TE}}$4, projected covariates $d_{\mathrm{TE}}$5, static features $d_{\mathrm{TE}}$6.
• Encoder: stacked ResidualBlock MLPs operating on concatenated, flattened inputs:

$d_{\mathrm{TE}}$7

• Decoder: stacked ResidualBlocks to yield $d_{\mathrm{TE}}$8, reshaped and further processed per-horizon-step via a temporal decoder. Direct global linear residual is added, i.e.,

$d_{\mathrm{TE}}$9

Empirical variants adapt output heads for quantile regression and value function approximation (Das et al., 2023, Chen et al., 17 Jan 2025).

3. Integration with Machine Learning and Control Paradigms

Sequence Modeling and Forecasting

TiDE can replace Transformer-based or recurrent forecasters where long context, non-linear covariate interactions, or missingness are present. Models have demonstrated:

• Linear scaling with sequence length and horizon, maintaining $TE(t)$0 time and space complexity.
• Ability to encode covariates and static attributes both in projection and decoder fusions.
• Near-optimal theoretical prediction error (in the linear submodel) for linear dynamical systems, with memory length $TE(t)$1 yielding $TE(t)$2-close performance to LDS-optimal predictors (Das et al., 2023).

Reinforcement Learning and RL-Control

In reinforcement learning for continuous state/action domains:

• TiDE is used as state encoder, mapping history to a latent vector $TE(t)$3 in DDPG, with no changes to standard actor/critic updates aside from this replacement.
• Empirical results in asset allocation indicate higher Sharpe ratios (1.13 vs. 0.95) and realized portfolio values compared to both Q-learning and passive strategies (Liu et al., 12 Aug 2025).

Model Predictive Control (MPC) and Digital Twins

TiDE acts as a one-shot, multi-step surrogate model:

• Simultaneous prediction across a horizon avoids recursive calls, yielding sub-second solve times for 50-step lookaheads.
• Integration with quantile regression enables robust, chance-constrained control, where quantile-based tightening of safety constraints reduces conservatism in robust MPC (failure rates: quantile-MPC 5.8%, tube-MPC 6.2%, nominal MPC 56.3%) (Chen et al., 17 Jan 2025).
• Demonstrated in manufacturing, TiDE-based MPC achieves precise setpoint tracking and safety constraint satisfaction not feasible with classical proxies or one-step RNNs (Chen et al., 10 Jan 2025).

4. Empirical Performance and Evaluation

Performance on Standard Forecasting Tasks

Empirical benchmarks include:

• On Electricity/Traffic/ETT datasets, TiDE achieves mean squared error (MSE) lower than Transformer (PatchTST) and autocorrelation-based models (DLinear). For example, on Traffic/H=720: TiDE 0.3868 vs PatchTST 0.4326 vs DLinear 0.4660 (Das et al., 2023).
• Ablation results indicate: removing skips degrades MSE by 3–7%; omission of temporal decoder leads to poor adaptation to rapid covariate-driven change.
• For high-capacity forecasting (M5), TiDE achieves WRMSSE 0.611±0.009 (lower than DeepAR’s 0.789±0.025 and PatchTST’s 0.976±0.014) (Das et al., 2023).

Real-Time MPC Applications

• TiDE-based robust MPC achieves sub-second (0.18s–0.28s) compute time per control step in high-dimensional horizon tasks where sequential RNNs would be at least 50× slower (Chen et al., 10 Jan 2025).
• For directed energy deposition additive manufacturing,