Boosted Direct Output (BDO) Forecasting

• BDO is a forecasting strategy in LTSF that combines AR's causality with DO's efficiency through a hierarchical, multi-stage procedure.
• It employs recursive sub-forecasts with dedicated time and frequency loss terms and EMA-based parameter smoothing to mitigate error accumulation.
• Empirical results demonstrate that BDO outperforms traditional AR and DO methods in benchmarks like weather, traffic, and energy datasets.

Searching arXiv for the cited paper and closely related LTSF context. Boosted Direct Output (BDO) is a forecasting strategy for Long-term Time Series Forecasting (LTSF) introduced in "ReNF: Rethinking the Design Space of Neural Long-Term Time Series Forecasters" (Lu et al., 30 Sep 2025). It is designed to combine the advantages of Auto-Regressive (AR) and Direct Output (DO) forecasting: AR instills causality by recursively predicting the next time step, but accumulates errors over long horizons, whereas DO predicts the entire forecasting window in a single forward pass, avoiding recursive error accumulation but neglecting causality. BDO addresses this tension through a hierarchical multi-step procedure that recursively extends a forecast in stages, conditions later stages on prior sub-forecasts, and applies supervision at each stage. In the paper’s formulation, BDO is paired with a Multiple Neural Forecasting Theorem (MNFT), a hierarchical loss over time and frequency domains, and parameter smoothing via Exponential Moving Average (EMA) (Lu et al., 30 Sep 2025).

1. Conceptual setting and motivation

The paper situates BDO within a critique of Neural Forecasters (NFs) for LTSF: progress has been hampered by an overemphasis on architectural complexity at the expense of fundamental forecasting principles (Lu et al., 30 Sep 2025). Within that framing, the central design question is not only how expressive the forecasting architecture is, but how the forecasting procedure itself organizes causality, horizon decomposition, and supervision.

The two baseline paradigms are presented as follows. Auto-Regressive (AR/AR(1)) predicts the next time step recursively, using previous predictions as inputs. This instills causality but accumulates errors over long-term horizons. Direct Output (DO) predicts the entire forecasting window in a single forward pass, which avoids recursive error accumulation but does not explicitly account for sequential relationships within the predicted future. The paper characterizes DO’s lack of causality as "anti-intuitive" and argues that it limits the model’s ability to build granular, robust, long-term forecasts (Lu et al., 30 Sep 2025).

BDO is motivated as an overview of these two regimes. It leverages the robustness and parallelism of DO while reinstating the sequential, causal structure associated with AR. The stated goal is to combine AR’s causality with DO’s efficiency and robustness without inheriting the full error propagation of step-by-step recursion. This suggests that the forecasting strategy, rather than architectural novelty alone, is treated as a primary design axis in LTSF.

2. Theoretical basis: Multiple Neural Forecasting Theorem

The theoretical foundation for BDO is the Multiple Neural Forecasting Theorem (MNFT) (Lu et al., 30 Sep 2025). The theorem addresses the setting in which multiple independent forecasts can be generated, for example via different parameters or random states. Its key insight is that there exists a post-combination function $g_c$ which, if accessible, can synthesize a better-than-individual forecast by leveraging the diversity among multiple outputs.

Given candidate forecasts $\{\hat{Y}_f^{(i)}\}_{i=1}^c$, the post-combined forecast is written as

$\hat{Y}_{pc}=g_c(\{\hat{Y}^{(i)}_f\}_{i=1}^c)$

The theorem further states that, under mild assumptions, the combined forecast error can be bounded and reduced by increasing the number of candidate forecasts (Lu et al., 30 Sep 2025). The practical obstacle is that the optimal $g_c$ typically requires future knowledge. The paper therefore does not attempt to realize an oracle post-combination directly; instead, it proposes a practical scheme in which the model implicitly learns to benefit from multiple staged forecasts.

In that sense, BDO can be read as an operational approximation to the theorem’s premise. Rather than explicitly constructing an oracle combination module, it recursively generates intermediate forecasts and feeds them back into subsequent forecasting stages. A plausible implication is that BDO functions as an internalized forecast-combination mechanism, using staged prediction and shared representation learning in place of explicit oracle access.

3. Forecasting mechanism and recursive structure

BDO is defined over a history $T_x=\{x_1,\dots,x_h\}$ and a target forecasting window $T_y=\{y_1,\dots,y_L\}$ (Lu et al., 30 Sep 2025). Instead of forecasting one step at a time as in AR, or all $L$ future steps at once as in DO, BDO recursively extends the forecast in stages. At stage $n$, the model predicts up to a longer horizon $h_n$ by taking as input both the original history and its own forecasts from prior stages:

$$\hat{T}_y\{1:h_n\} = \mathrm{NF} \left( [T_x, \hat{T}_y\{1:h_{n-1}\}] \right), \quad n=1,2,...,N$$

where $\{\hat{Y}_f^{(i)}\}_{i=1}^c$0 denotes temporal concatenation, $\{\hat{Y}_f^{(i)}\}_{i=1}^c$1, and $\{\hat{Y}_f^{(i)}\}_{i=1}^c$2 (Lu et al., 30 Sep 2025). In practice, $\{\hat{Y}_f^{(i)}\}_{i=1}^c$3 is split into $\{\hat{Y}_f^{(i)}\}_{i=1}^c$4 even or configurable segments.

This recursion defines the core distinction between BDO and the two standard baselines. AR predicts one step at a time and feeds each prediction as input; error accumulates across steps. DO performs a single-shot prediction of all future steps, which avoids recursive accumulation but does not enforce internal causality among outputs. BDO instead yields a hierarchical multi-step procedure: intermediate sub-forecasts are produced, then reused as conditioning variables for deeper stages. The result is a chunked recursive forecast that preserves sequential structure without reducing forecasting to strictly one-step recursion (Lu et al., 30 Sep 2025).

The paper attributes several benefits to this mechanism: causality, reduced error accumulation relative to pure AR, hierarchical curriculum, scalability through adjustable depth, and implicit post-combination. The claim about scalability is especially specific: deeper BDO architectures empirically yield better performance, unlike simple DO where depth brings little benefit (Lu et al., 30 Sep 2025). This suggests that in BDO, additional depth has procedural meaning because each block expands and refines the forecast horizon rather than merely increasing architectural capacity.

4. Hierarchical supervision, objective function, and stabilization

A central component of BDO is hierarchical supervision (Lu et al., 30 Sep 2025). Each stage outputs a sub-forecast and receives its own loss term, with earlier, shorter horizons weighted more heavily. The paper explicitly describes this as having a curriculum learning effect. Because supervision is applied at multiple levels of the forecast hierarchy, the model receives richer feedback than in a one-shot DO regime.

The total loss is defined as

$\{\hat{Y}_f^{(i)}\}_{i=1}^c$5

where $\{\hat{Y}_f^{(i)}\}_{i=1}^c$6 is the forecast at stage $\{\hat{Y}_f^{(i)}\}_{i=1}^c$7 for a partial future, $\{\hat{Y}_f^{(i)}\}_{i=1}^c$8 is the corresponding ground truth, $\{\hat{Y}_f^{(i)}\}_{i=1}^c$9 denotes the Discrete Fourier Transform, and $\hat{Y}_{pc}=g_c(\{\hat{Y}^{(i)}_f\}_{i=1}^c)$0 are tuning coefficients (Lu et al., 30 Sep 2025). The loss therefore combines a time-domain $\hat{Y}_{pc}=g_c(\{\hat{Y}^{(i)}_f\}_{i=1}^c)$1 term with a frequency-domain $\hat{Y}_{pc}=g_c(\{\hat{Y}^{(i)}_f\}_{i=1}^c)$2 term, and scales the contribution of each stage by $\hat{Y}_{pc}=g_c(\{\hat{Y}^{(i)}_f\}_{i=1}^c)$3.

This learning objective serves two roles. First, it increases supervision density by attaching losses to intermediate sub-forecasts rather than only to the terminal horizon. Second, it aligns with the MNFT perspective by treating intermediate outputs as meaningful forecast candidates rather than disposable hidden states. The paper states that this multi-level, hierarchical loss uses more supervision and provides richer feedback to each sub-forecast (Lu et al., 30 Sep 2025).

Training is further stabilized by smoothly tracking the model’s parameters. In implementation terms, the method uses Exponential Moving Average (EMA) of parameters for evaluation, smoothing out instabilities common in time series training (Lu et al., 30 Sep 2025). This is presented as a training stabilization mechanism rather than a separate forecasting model.

5. Implementation pattern and relation to AR and DO

The implementation described for BDO is intentionally simple (Lu et al., 30 Sep 2025). The method can be used with simple MLPs or more complex architectures; each BDO block gets a linear head for its sub-forecast, blocks are stacked, and each block outputs a forecast for a growing segment of the prediction window. At each stage, the input pipeline concatenates the original history with the outputs of all previous sub-forecasts. The hybrid time-and-frequency loss is then applied at each stage, weighted in favor of shorter, easier horizons.

The paper’s comparative summary of AR, DO, and BDO is captured in the following scheme:

Property	AR	DO	BDO
Causality	Yes	No	Yes (via recursion)
Error Accumulation	High	Low	Low–Moderate (chunked)
Output Granularity	1-step	Multi-step	Multi-step, hierarchical
Supervision Usage	Each step	Only once	Multiple, at each stage
Empirical Performance	Weak (LTSF)	Good	Best (in paper results)

This comparison clarifies that BDO is not merely an interpolation between AR and DO at the level of horizon length. Its defining property is hierarchical recursion over multi-step sub-forecasts. That structure creates a causal dependency among outputs while preserving chunked prediction and repeated supervision. A common misconception in this design space is that avoiding one-step recursion is equivalent to abandoning causality; the BDO formulation explicitly rejects that equivalence by conditioning each stage on preceding forecast segments (Lu et al., 30 Sep 2025).

6. Empirical results, ablations, and bound verification

The empirical claim of the paper is that these principled changes enable a simple MLP to achieve state-of-the-art performance, outperforming recent, complex models in nearly all cases (Lu et al., 30 Sep 2025). The reported benchmarks include ETT, Traffic, Weather, Solar, and related datasets. Specific examples given in the detailed summary are Weather with ReNF (BDO) MSE $\hat{Y}_{pc}=g_c(\{\hat{Y}^{(i)}_f\}_{i=1}^c)$4, TimeBridge MSE $\hat{Y}_{pc}=g_c(\{\hat{Y}^{(i)}_f\}_{i=1}^c)$5, and DUET MSE $\hat{Y}_{pc}=g_c(\{\hat{Y}^{(i)}_f\}_{i=1}^c)$6; Traffic with ReNF MSE $\hat{Y}_{pc}=g_c(\{\hat{Y}^{(i)}_f\}_{i=1}^c)$7, TimeBridge MSE $\hat{Y}_{pc}=g_c(\{\hat{Y}^{(i)}_f\}_{i=1}^c)$8, and DUET MSE $\hat{Y}_{pc}=g_c(\{\hat{Y}^{(i)}_f\}_{i=1}^c)$9; and ETTm2 with ReNF MSE $g_c$0, TimeBridge MSE $g_c$1, and DUET MSE $g_c$2 (Lu et al., 30 Sep 2025). The paper further states that the method often operates with far less computational overhead.

Ablation results emphasize the role of depth. Increasing sub-forecast stages, described as BDO depth, improved performance, while adding depth to DO led to no gain or even worse results (Lu et al., 30 Sep 2025). This empirical contrast is important because it supports the claim that BDO depth is functionally meaningful rather than simply an increase in parameterized capacity.

The paper also reports an Empirical Bound (MNFT Oracle) experiment. By simulating an oracle post-combination among multiple sub-forecasts and choosing the best prediction for each step, it demonstrates that BDO narrows—but does not close—the gap with the empirical error bound established by MNFT (Lu et al., 30 Sep 2025). The abstract summarizes this as empirically verifying the theorem, establishing a dynamic performance bound, and identifying promising directions for future research. This suggests that the current BDO mechanism is an approximation to a stronger, not directly realizable forecast-combination process, leaving open the question of how to learn better post-combination strategies without oracle access.

7. Significance within LTSF design space

Within the paper’s framing, BDO represents a return to first principles in LTSF (Lu et al., 30 Sep 2025). The emphasis is on redesigning the forecasting paradigm rather than increasing architectural complexity. By combining staged recursion, hierarchical supervision, and parameter smoothing, the method redefines how an NF should construct a long-horizon forecast.

The broader significance claimed by the paper is twofold. First, BDO is described as simple and model-agnostic: even MLPs perform as well as or better than transformer-based state-of-the-art models when BDO is applied appropriately (Lu et al., 30 Sep 2025). Second, the method is explicitly tied to a theorem about multiple forecasts and post-combination, giving the design a theoretical rationale rather than presenting it solely as an empirical heuristic.

The limitations are also visible in the same framing. The optimal post-combination function $g_c$3 typically requires future knowledge, so BDO only implicitly approximates the theorem’s ideal. The empirical oracle bound confirms remaining headroom between current staged forecasting and the best achievable combination over multiple candidate outputs (Lu et al., 30 Sep 2025). A plausible implication is that future work may focus less on increasingly complex backbones and more on richer internal combination strategies, improved staged forecast interaction, and tighter approximations to the MNFT post-combination regime.