Error-Adaptive Ensemble Model for Robust

Demand Forecasting in Supply Chain Management

S. Sreenivasula Reddy Dr. N. Gopinath
Department of Computational Intelligence Department of Computational Intelligence
SRM Institute of Science and Technology SRM Institute of Science and Technology
Kattankulathur, Chennai, India Kattankulathur, Chennai, India
ss3708@[Link] gopinatn@[Link]
Abstract— Demand forecasting is a crucial component of
supply chain management; however, the efficacy of traditional For a long time, most forecasting toolboxes used more
statistical models and single-architecture deep learning models
is significantly compromised when subjected to real-world, traditional stats-based methods like ARIMA [5] or
volatile, and non-stationary demand patterns. This paper exponential smoothing. These models are easy to understand
introduces an innovative error-adaptive ensemble architecture and work well when the data shows stable, even linear
that integrates the Patch Time Series Transformer (PatchTST) relationships[6], but they don't work well with the nonlinear
with gradient boosting machines, resulting in markedly
improved demand forecasting accuracy, even amidst varying complexities and interdependencies that modern supply
degrees of volatility. Our proposed architecture integrates the chains create[7].
self-attention mechanisms of PatchTST on temporal patches to
identify long-range dependencies and underlying trends in the For example, the person who made the decomposed
data, and employs XGBoost with residual learning to
dynamically rectify prediction errors through gradient boosting models like Prophet[8] has recently had great success with
on feature-engineered residuals. A validation-driven adaptive dealing with changes in trends and seasonality effects.
weighting strategy is used to change the relative weight of the Nonetheless, despite the advancement and proliferation of
two parts in real time. This lets the ensemble work well even decomposed models for addressing time series issues, the
when demand patterns are very different. We have fully tested
our proposed ensemble architecture on 127 product-location capture of intricate temporal dependencies continues to pose
time series from a real-world supply chain dataset. It has a a significant challenge. Deep learning provided an
mean absolute percentage error (MAPE) of 52.41%, which is a alternative through the advancement of the natural
big improvement of 36.3% over ARIMA (82.34%), 66.6% over extensions of Long Short-Term Memory (LSTM) networks
Prophet (156.78%), 29.7% over LSTM (74.56%), and 26.4%
over XGBoost (71.23%). Volatility regime analysis demonstrates
[9] to address nonlinearities in temporal dependencies. But
that our proposed ensemble outperforms even in scenarios of neural networks might have trouble with overfitting, tuning
low-volatility (41.23% MAPE), medium-volatility (55.78% hyperparameters, and changes in the training data's
MAPE), and high-volatility (60.45% MAPE) demand patterns. distribution [10].
It achieves a R² value of 0.46, representing a significant
enhancement over baseline models (R² values ranging from -
0.45 to 0.28), and establishes a significant predictive relationship Recently, attention-based architectures have changed the
in highly volatile data where traditional models fail to exceed field of time-series models. Transformers [11], which were
the accuracy of naive forecasting benchmarks. first used for language models, have now been used to
predict time series. PatchTST [12] and other methods break
Keywords: Time series analysis, adaptive weighting, gradient a time series into patches, which makes it easier to see short-
boosting, transformer models, ensemble learning, demand
forecasting, supply chain management, and intermittent term and long-term patterns in the data, even with self-
demand attention.

B. Research Gap
I. INTRODUCTION
Even with these improvements, there are still some big
A. Background and Motivation problems that need to be fixed:
In today's business world, supply chain management is a  Limitations of single models: Individual models,
moving target; it is becoming more complicated, unstable, whether they are statistical or neural, work well in
and unpredictable [1], [2]. It's clear that demand is the most some situations but not in others[13].
important part of good inventory management, production  Error systematicity: Deep learning models often
planning, and logistics management[3]. Real-world demand make systematic mistakes when predicting things that
is messy, though. It doesn't always stay the same; it has simple ensembling can't fix [14].
several seasonal cycles; promotional activity can create extra  Sensitivity to volatility: Most forecasting methods
lift; and it can change quickly in response to outside don't work as well on demand streams with a lot of
shocks[4]. volatility, which is where accurate predictions are most
important [15].
  Static ensembling: Traditional ensemble methods use piecewise-linear dynamics. More complex nonlinear
fixed weighting schemes that don't change when the interactions can make them less useful[20].
accuracy of the forecast changes over time [16].
C. Research Contribution B. Deep Learning for Time Series
This study tackles these issues by suggesting a new The application of deep learning for time series prediction
ensemble architecture that is aware of errors and uses the has evolved over the years with the development of new
best parts of attention-based deep learning and gradient architectures:
boosting. At its core, PatchTST has two parts: it finds hidden Recurrent Neural Networks: Standard RNNs had a
patterns, and a residual learning module controlled by problem with vanishing gradients, but LSTM networks [9]
XGBoost is in charge of finding and fixing systematic errors. and Gated Recurrent Units [21] solved it. This made it
This is different from traditional ensemble methods that just possible to learn long-range relationships. DeepAR [22]
average predictions. We use a validation-adaptive weighting improved LSTMs by adding probabilistic forecasting and
mechanism that changes the relative weight of each autoregressive likelihood modeling.
component in real time to keep our system working well
even when the market changes. We test the proposed method
Convolutional Architectures: Temporal Convolutional
on 127 real-world product-location time series that cover a
Networks [23] used dilated causal convolutions to get large
wide range of volatility conditions. The final solution is built
receptive fields while still being fast to compute. WaveNet
to be able to handle the amount of computation needed to be
[24] demonstrated the efficacy of stacked dilated
used in a real-world supply chain forecasting system.
convolutions in generating sequential data.
The rest paper is structured as follows: Section II reviews
related work in time series forecasting and ensemble Models Based on Attention: The Transformer architecture
methods. Section III presents the proposed error-adaptive [11] changed how sequences are modeled by using self-
ensemble framework, including mathematical formulations attention processes that find global dependencies in a
and algorithmic details. Section IV describes the consistent amount of time. Temporal Fusion Transformers
experimental methodology, dataset characteristics, and [25] changed transformers so that they could be used for
evaluation protocols. Section V presents comprehensive multi-horizon forecasting by adding attention patterns that
results and analysis. Section VI discusses theoretical can be understood. ProbSparse self-attention was created by
insights, practical implications, and limitations. Section VII Informer [26] to make it easier for computers to predict long
concludes with future research directions. sequences.
II. RELATED WORK
Right now, PatchTST [12] is the best patch-based
A. Classical Time Series Forecasting transformer. It splits time series into pieces that can be used
For a long time, the best way to predict time series has as input tokens for transformer encoders. This method
been to use classic statistical methods. The Box-Jenkins shortens the length of the sequence, lets you learn from other
methodology[5] produced the ARIMA models, which, in tasks, and improves both accuracy and speed compared to
theory, divide a series into three parts: autoregressive, pointwise transformer inputs.
integrated, and moving average. After that, SARIMA[17]
came along, which combined seasonal patterns by using C. Ensemble Methods in Forecasting
seasonal differencing and seasonal AR/MA terms.
Using variety and aggregation, ensemble learning
combines different models to get better results [20].
Exponential Smoothing uses a different way to weight
Ensemble methods have consistently surpassed alternative
past observations to make predictions. There are different
techniques in time series forecasting
types of smoothing, such as simple, double, and triple
smoothing, depending on whether the series has level, trend,
and seasonality[6][18]. The state-space perspective Simple Averaging for putting equal weight on many
illustrated exponential smoothing within a cohesive forecasts has been shown to lower prediction variance and
statistical framework. make predictions more reliable [28], [29]. Weighted
Prophet[8] recently came up with an additive decomposition Ensembles for systems that weight models based on how
that has piecewise trends, multiple seasonalities, and holiday well they work give more weight to models that are more
effects. It is made for business time series that have a lot of accurate [30]. It is still hard to figure out the best weight, and
seasonality and change regimes often. there are many ways to do it, such as inverse-error weighting
and regression-based algorithms [31]. Stacked generalization
[32] teaches meta-models how to use advanced aggregation
These classical methods are simple to understand and
functions to combine base predictions. Recent studies have
quick to calculate, but they only work well for linear or
employed neural networks as meta-learners [14].
 |xvali | = 0.15 · Ti, and
Puzzles for Merging Predictions of empirical research has |xtesti | = 0.15 · Ti.
demonstrated scenarios where simple averaging outperforms
optimal weighted combinations [33], highlighting the The forecasting objective is to learn a function f : RL → RH
importance of robustness over theoretical optimality. that maps a lookback window of length L to a prediction

D. Gradient Boosting for Time Series horizon of length H, minimizing expected forecasting error:
Gradient boosting machines [34] have been very good at
using sequential ensemble learning to predict structured data.
f∗¿ argmi n {f ∈ F } E ( 2)
XGBoost [35] and LightGBM [36] made scaled versions that { (x , y ) D} [ L (f (x { t−L :t} ) , y {t +1:t +H } ) ]

used regularization and made it easier to build trees.

Most of the time, gradient boosting has been used for time where L is a loss function (usually mean squared error for
series forecasting by making new features from lagged data, training models) and F is the space of hypotheses for
rolling statistics, and temporal encodings [37]. Standalone forecasting models.
gradient boosting models, however, struggle with long-range B. PatchTST Architecture
dependencies and seasonal patterns without extensive human PatchTST [12] is the first part of our ensemble. It is a
feature engineering. cutting-edge transformer-based forecasting model that works
E. Research Positioning on patched time series segments instead of single time
points.
The suggested method includes several important new
ideas that make it different from other hybrid forecasting 1) Patch Embedding: We first divide an input time
methods. While previous methods have looked at combining series x = (x1,x2,...,xL) of length L into P-sized patches that
neural networks and boosting algorithms, the proposed don't overlap:
method is different because it uses PatchTST to better model
L
the data's temporal aspects to find long-term trends and Patc h j =( x {( j−1) P+1 } , x {( j−1) P +2 } , … , x { jP }) , j=1 ,2 , … , ⌊ ⌋ ( 3)
seasonality, and XGBoost to learn the complex patterns of P
the residuals that are left after the deep learning predictions. Each patch is then linearly projected to a d-dimensional
The proposed approach is different from other methods that embedding space:
just combine the predictions of different models. Instead, it
uses a clear method for residual modeling, where XGBoost z j =W e · Patc h j +b e ( 4 )
where We ∈ Rd×P and be ∈ Rd are learnable parameters.
is trained on the residual errors made by PatchTST. This lets
errors in forecasting be corrected systematically instead of
just averaging the predictions. The proposed method also
uses an adaptive weighting scheme that changes based on 2) Positional Encoding: To preserve temporal ordering
the data's characteristics without needing any tuning. This information, we add learnable positional embeddings:
makes it more stable for different demand patterns. Finally,
the proposed method is made for supply chain forecasting, ~z =z + p ( 5 )
j j j

where pj ∈ Rd is the positional embedding for patch j.

which has to deal with the problems of demand patterns that
are irregular, changeable, and not stationary, which is very
much like what happens in real-world supply chains. 3) Transformer Encoder: The sequence of patch
{~z j }{{ j=1} } where N p=⌊ L/ P ⌋
N
embeddings p
is processed
III. METHODOLOGY
through M transformer encoder layers. Each layer m consists
A. Problem Formulation of multihead self-attention and position-wise feed-forward
{N }
Let D=\{ ( x i , y i ) \}{i=1 } be a set of N univariate time operations:
series. Each series x i=(x i , 1 , x i, 2 ,... , x i ,T i) shows Multi-Head Self-Attention:
O
historical demand observations for a certain productlocation MultiHead ( Z ) =Concat ( hea d 1 , … ,hea d K ) W ( 6 )
combination, and y i=( y i ,1 , y i , 2 ,... , y i , H) shows
the future demand values over a forecasting horizon H.
where each attention head is computed as:
We divide the data into three nonoverlapping sets for each
time series i, as is standard in machine learning
hea d k =Attention ( Z W Qk , Z W Kk , Z W Vk ) ( 7 )
xi = xtraini ∪ xvali ∪ xtesti (1)

where: |xtraini| = 0.70 · Ti,

 ( )
QK
T 2) Feature Engineering for Residuals: We construct a
Attention ( Q , K ,V ) =softmax V ( 8) feature matrix for residual prediction using lagged residual
√ dk

[ ]
values and temporal encodings:
T
r t −1 ,
Here, WkQ,WkK,WkV ∈ Rd×dk are learnable projection
r t−7 ,
matrices, and K is the number of attention heads. Position- r t −14 ,
Wise Feed-Forward Network: f t= mean ( r t−1 : t−1 ) , ( 15 )
mean ( r t−7 : t−1 ) ,
FFN ( z )=max ( 0 , z W 1 +b 1 ) W 2+ b2 ( 9 )
dow ( t ) ,
Each sub-layer employs residual connections and layer month ( t )
normalization:
where dow(t) and month(t) represent day-of-week and
Z =LayerNorm ( Z + MultiHead ( Z ) ) ( 10 )
m m −1 m−1
month encodings, respectively.
3) XGBoost Training: We train an XGBoost regressor
[6] to predict residuals:
Z =LayerNorm ( Z + FFN ( Z ) ) (11 )
m m m

K
r^ t =Σ k=1 f k ( f t ) ( 16 )
4) Forecasting Head: The final transformer output is
flattened and passed through a linear projection to generate where fk represents the k-th regression tree in the
forecasts: ensemble, learned through gradient boosting:

ŷ PatchTST =Flatten ( Z M ) W out +b out ( 12 ) N

f k =arg mi n f Σ i=1 L ( r i , Σ Kj=1
train −1
f ( f i ) + f ( f i ) )+ Ω ( f ) ( 17 )
where W out ∈ R(N p · d)× H and b out ∈ R H .
5) Training Objective: PatchTST is trained to Here, L is the squared loss function, and Ω(f) is a
minimize mean squared error: regularization term:

LPatchTST = ( N·H1 ) Σ N
i=1 Σ Hh=1 ( y i ,h − ŷ i ,h )2 ( 13 )
1 T 2
Ω ( f )=γT + λ Σ j=1 w j ( 18 )
2
We use the Adam optimizer [22] with a learning rate of η =
0.001 and train for E = 30 epochs, stopping early if the
validation loss goes down. where T is the number of leaves in tree f, wj is the score
assigned to leaf j, and γ,λ are regularization hyperparameters.
C. Residual XGBoost Correction We use the following XGBoost configuration:
The main new thing about our framework is that it uses • Number of estimators: K = 100
gradient boosting to model PatchTST's prediction errors
• Maximum tree depth: dmax = 3
directly. This residual learning method is based on the fact
that deep learning models often make the same mistakes • Learning rate: ηXGB = 0.05
over and over again, even when the time period changes. • Subsample ratio: ρ = 0.8
1) Residual Calculation: We calculate residuals on the • Column subsample ratio: ρcol = 0.8
same training data after training PatchTST on x train.:
D. Adaptive Weighting Mechanism
PatchTST Our framework uses a validation-driven adaptive
r t =x t – x^ t ( 14 )
weighting scheme that dynamically balances the
contributions of PatchTST and the combined
where xˆPatchTSTt is the PatchTST prediction at time t during PatchTST+XGBoost prediction. This is different from
one-step-ahead forecasting. traditional ensemble methods that use fixed weights.
1) Component Performance Evaluation: We calculate
forecasting errors for two configurations on the validation
set xval:
Configuration 1 - PatchTST Only:

patch=RMSE ( y i , ŷi
ε (i) ) ( 19 )
val PatchTST

Configuration 2 - PatchTST + XGBoost:

ensemble PatchTST XGBoost
ŷi = ŷi + r^ i ( 20 )

ensemble =RMSE ( y i , ŷ i
ε (i) ) ( 21 )
val ensemble

2) Inverse Error Weighting: We compute inverse-error

weights:
(i) 1 (i) 1
w patch= (i)
, w ensemble= (i )
(22)
ε patch +δ ε ensemble +δ
where δ = 10−6 is a small constant to prevent division by
zero.
3) Weight Normalization: Weights are normalized to
sum to unity

(i)
(i) w patch
α patch = ( 23 )
( w(i)patch +w(i)ensemble)
(i)
(i) w ensemble
α ensemble = ( 24 )
( w(i)patch + w(iensemble
)
)
4) Final Prediction: The final forecast combines both
components using adaptive weights: The computational complexity of our framework consists
of three main components:
 PatchTST Training: The transformer encoder has

is the number of layers, Np = ⌊L/P⌋ is the number of

complexity per forward pass, where M
Ŷ final =α (i) ensemble · ( ŷ i
+ α (i) + r^ iXGBoost ) ( 25 )
PatchTST PatchTST
i patch · ŷ i
patches, and d is the embedding dimension.
This formulation can be simplified as:  XGBoost Training: Gradient boosting tree
construction has complexity O(K · df · Ntrain · dmax),
final PatchTST (i) XGBoost where K is the number of trees, df is the feature
Ŷi = ŷi +α ensemble · r^ i ( 26 ) dimension, Ntrain is the training set size, and dmax is
maximum tree depth.
(i)  Inference: Prediction for a single time series requires
Thus, the adaptive weight α ensemble controls the magnitude
2
of residual correction applied to each time series, with higher one PatchTST forward pass O(N p · d) and one
weights assigned when the XGBoost component XGBoost prediction (O(K · dmax)).
demonstrably improves validation performance. For large-scale deployment on N time series, the total
E. Algorithmic Summary training complexity is
The complete training and inference procedure is O(N·E· ( N 2p · d + K· d f ·T i · d · d max ) ), where E is the
presented in Algorithm 1. number of training epochs.

IV. EXPERIMENTAL METHODOLOGY

A. Dataset Description
We apply the proposed framework to a retailer's supply
chain dataset that displays daily sales of various products
across multiple channels from January 2013 to June 2016.
We added up the sales for each pair of products and locations We configure Prophet with yearly seasonality enabled
every day. Days when there were no sales were counted as (Fourier order = 10), weekly seasonality enabled (Fourier
having no sales. Outliers were kept, but they were marked if order = 3), and changepoint prior scale of 0.05.
they were more than five standard deviations away from the 3) LSTM: We implement a single-layer LSTM
30-day moving average. There were calendar variables, like network for one-step-ahead forecasting:
the day of the week, the month, and the holidays. The data • Input: Lookback window of 7 days
was combined to make 127 product-location demand time • LSTM layer: 16 hidden units
series. These were then modeled as separate time series with • Dense output layer: 1 unit
different patterns. There are between 100 and 1,300 days in • Loss function: Mean Squared Error
the time series, with 425 days being the most common. We • Optimizer: Adam with learning rate η = 0.001 •
divided each time series into three groups: training (70%), Epochs: 50 with early stopping
validation (15%), and testing (15%). There were no two sets
that were the same. The input data for deep learning 4) XGBoost (Standalone): As an additional baseline,
architectures such as PatchTST and LSTM was normalized we train XGBoost directly on the forecasting task with
to a range of [0,1]. feature engineering:
f t= [ xt −1 , x t−7 , x t−14 , mean ( x t −3 : t−1 ) , mean ( x t −7 :t −1) , dow ( t ) , m
~ ( x t −min ( x ) )train

( 27 )
xt = Hyperparameters match the residual XGBoost component
( max ( x train )−min ( x train ) ) configuration.
D. Evaluation Metrics
Traditional statistical models (ARIMA, Prophet) and We assess forecasting performance using four
XGBoost operate on the original scale without complementary metrics:
normalization.
1) Mean Absolute Error (MAE):
B. Volatility Regime Classification
To analyze model robustness across different demand 1
patterns, we categorize each time series into volatility
(i)
MA E = Σ t ∈T | y(i)t − ŷ (i)t |( 32 )
|T |
test
test
regimes based on the coefficient of variation (CV):
2) Root Mean Squared Error (RMSE):
σ ( x train
i )
C V (i)= (28 )
μ ( xi )
√[
train
(i) 1
RMS E =
|T test|
Volatility Categories:
• Low Volatility: CV < 0.20 (stable, predictable demand)
3) Mean Absolute Percentage Error (MAPE):
• Medium Volatility: 0.20 ≤ CV ≤ 0.35 (moderate
variation)

| |
i i

(| | )
• High Volatility: CV > 0.35 (highly erratic, spiky i 100 yt − ŷt
demand) MAP E = test
Σ {t ∈T test
} i
( 34 )
C. Baseline Models
T yt + ε
We compare the proposed error-adaptive ensemble against where ϵ = 0.1 prevents division by
four established forecasting approaches: zero.
1) ARIMA: We implement seasonal ARIMA with
automatic order selection: 4) Coefficient of Determination (R2):
2
Φ P ( Bs ) φ p ( B ) ∇ sD ∇d x t =ΘQ ( B s ) θq ( B ) ε t ( 29 ) 2 (i )
Σ t ∈T test ( y it − ŷ it )
R =1− 2
( 35 )
( y it − ȳ i )
(AIC) with p,d,q ∈ {0,1,2} and seasonal order (P,D,Q,s)
Order selection uses the Akaike Information Criterion Σ t ∈T test

with P,D,Q ∈ {0,1} and s = 7.

Aggregation Across Time Series: For overall model
comparison, we compute the arithmetic mean of each
2) Prophet: Prophet [41] decomposes time series into
metric across all 127 time series:
trend, seasonality, and holiday components:
1 (i )
y t =g ( t ) + s ( t )+ h ( t ) + ε t ( 30 ) Metric= Metric (36)
N
E. Statistical Significance Testing linear models are better for figuring out how demand
To assess whether performance differences between changes quickly. XGBoost is still a strong competitor
models are statistically significant, we employ the Wilcoxon because it can do good feature engineering. This shows that
signedrank test [45], a non-parametric paired test. We report tree-based models are better for forecasting graphical time
p-values and consider differences significant at α = 0.05 series.
level with Bonferroni correction.
B. Volatility Regime Analysis
FIGURE II
V. RESULTS AND ANALYSIS MAPE (%) BY VOLATILITY REGIME
A. Overall Performance Comparison
FIGURE I presents the aggregate forecasting performance
of all models across the 127 test time series.
FIGURE I
OVERALL FORECASTING PERFORMANCE

FIGURE II decomposes MAPE performance by volatility

category.
The proposed ensemble model's performance is even
better because it has the lowest MAPE at all levels of
demand volatility. The suggested ensemble model works
much better than the current method, XGBoost, at all levels
The experiment's results show that the suggested ensemble of demand volatility. In a low-volatility environment with
model works well based on a number of different ways to stable time series, the proposed ensemble model gets a
measure it. The ensemble model's MAPE of 52.41% is MAPE of 41.23%, which is a big improvement over
26.4% better than the best baseline model, XGBoost XGBoost's MAPE of 58.67%, or almost 30%. The proposed
(71.23%). It also beats ARIMA (82.34%) by 36.3% and ensemble model achieves a MAPE of 55.78% compared to
LSTM (74.56%) by 29.7%. The same thing happens with the XGBoost's MAPE of 75.89% in a medium volatility
absolute and squared error: MAE goes down from 12.87 to environment, where demand changes moderately. This is a
9.34, which is a 27.4% improvement over XGBoost, and relative improvement of 26.5%. It is important to note that
RMSE goes down from 18.45 to 13.67, which is a 25.9% this environment has the most data points, with 52 out of 127
improvement. The ensemble model's R² value is 0.46, which time series. This setting is also like a real-life situation in
is much higher than the baselines: ARIMA (0.12), LSTM supply chain forecasting. The suggested ensemble model
(0.24), and XGBoost (0.28). works well even in a very volatile environment where
demand changes a lot and is hard to predict. Its MAPE is
This R² value shows that the ensemble model accounts for 60.45%, which is a 26.6% relative improvement over
46% of the change in demand. This is a big deal because XGBoost's MAPE of 82.34%.
supply chain data is very unstable, and most traditional
models do worse than naive baselines. Prophet, on the other C. Adaptive Weight Analysis
hand, doesn't do well it has a MAPE of 156.78% and a R² of Illustrates the distribution of adaptive weights αensemble
-0.45. This is to be expected because its method of breaking computed on the validation set. Weight Statistics:
down demand into smaller parts doesn't work well for SKU-
level demand that isn't steady and doesn't happen all the • Mean: 0.58 ± 0.23
time. • Median: 0.62
• Interpretation Range: [0.18, 0.94]
Finally, when you look at the baselines, it's clear that By looking at the ensemble weights, you can see how the
LSTM is 9.5% better than ARIMA. This shows that non- proposed framework adapts to changing demand patterns on
its own. The weights are between 0.18 and 0.94, which where reliable medium- to long-term forecasts are very
means that there is a best mix of base forecasting and important.
residual correction for each time series. The median weight
VI. DISCUSSION
of 0.62 shows that residual correction is more important on
average. This means that the corrective part of the model is A. Theoretical Insights
more useful for improving accuracy for most time series. 1) Why Ensemble Outperforms Single Models: A bias-
There is a clear link between how weights are given and how variance-covariance analysis tells us how well our error-
volatile the demand time series is. The more volatile time adaptive ensemble works. Our method combines the best
series get higher weights (α > 0.7), and this link is parts of both solid and varied base learners. PatchTST learns
statistically significant with a Pearson r of 0.34 (p < 0.001). directly from raw time series data in a complete way, finding
This observation supports the idea that residual learning long-term dependencies and complicated patterns. This is
becomes more important when the base model's predictions why it is the best tool for modeling time. On the other hand,
are less accurate. XGBoost is great at finding structured residual patterns and
TABLE I approximating non-linear functions. Using different ways to
BIAS-VARIANCE DECOMPOSITION OF MSE
show things and different ways to learn can help lower error
Model Bias2 Variance Total MSE
covariance. Unlike traditional ensembles that just average
ARIMA 185.34 274.56 316.58 predictions, our method adds an error correction step: we
Prophet 892.34 691.23 1583.57 train XGBoost on PatchTST residuals. This targeted residual
LSTM 156.78 213.45 370.23 learning fixes the deep model's biases and persistent errors in
XGBoost 148.23 192.34 340.57 a systematic way, which makes it better at making
Ensemble 89.45 97.23 186.68
predictions that work in different demand situations.
D. Error Decomposition Analysis
2) Adaptive Weighting as Meta-Learning: The
Table I presents bias and variance estimates computed via validationdriven weighting scheme can be viewed as a form
bootstrapping. of metalearning [35] where the ensemble learns how to learn
1) Bias Reduction: The ensemble reduces bias 2 by 44.8% by observing model performance patterns. This approach
compared to baselines. shares conceptual similarities with mixture of experts [19]
2) Variance Reduction: The ensemble achieves 50.5% and online learning [36].
variance reduction.
3) Combined Effect: The simultaneous reduction in both
bias and variance yields a 48.0% total MSE reduction. B. Practical Implications
1) Operational Deployment: The proposed ensemble is
E. Forecast Horizon Sensitivity designed for integration into production forecasting systems:

are feasible given ∼24-minute training time for 127 series.

Batch Retraining on Daily or weekly retraining schedules
TABLE II
MAPE (%) BY FORECAST HORIZON Real-Time Inference is 4.1 ms per-series latency enables on-
Horizon ARIMA Prophet LSTM XGBoost Ensemble demand forecasting for tens of thousands of SKUs.
1-day 71.23 142.56 65.34 62.45 45.23 Confidence Intervals of PatchTST naturally extends to
3-day 78.45 152.78 70.12 67.89 49.67 probabilistic forecasting through quantile regression.
7-day 86.89 165.34 77.56 74.23 54.12 XGBoost’s interpretability feature importance provides
14-day 94.67 181.45 85.67 82.56 59.78 actionable insights.

Table II shows how well different prediction horizons 2) Business Value: The 28.5% MAPE reduction translates
work for forecasting and how strong the proposed ensemble to substantial operational benefits:
approach is. The ensemble consistently gets lower MAPE Inventory Cost Savings: Lower forecast errors reduce
values than the other models at all time horizons. This shows
[37], a 28.5% error reduction enables ∼20% reduction in
safety stock requirements. Using the newsvendor model
that it has a clear and lasting advantage in both short- and
long-term forecasting situations. As the horizon lengthens, safety stock while maintaining service levels.
all methods make more mistakes when trying to predict the Quantitative Example: For a mid-sized retailer with
future. This is a common problem in time series prediction, $100M annual inventory carrying costs, a 20% safety stock
but the ensemble shows a more gradual decline in reduction yields $20M annual savings.
performance. This means that the suggested method is better
at keeping useful time-based information and dealing with C. Limitations and Future Work
uncertainty over longer forecasting windows. This makes it The results are very good, but there are some clear
especially useful for real-world supply chain planning, problems that need to be fixed. The model is a simple one
that only looks at past demand. It doesn't take into account
other things that could make it work better, like changes in management. By synergistically combining PatchTST for
price, sales, the weather, or the economy. Second, the model trend modeling with XGBoost-based residual learning for
has negative R² values even when the conditions are very systematic error correction, and employing validation-driven
unstable. This shows that even the best machine learning adaptive weighting, the proposed method achieves
algorithms can fail when the conditions are very unclear. substantial performance improvements.
Third, the model's design to train a separate model for each The suggested ensemble works well in terms of both
time series is not scalable from a computational point of accuracy and speed. It gets a MAPE of 67.76%, which is
view, which makes it hard to use with large datasets, like 28.5% better than the best baseline. This means it can make
tens of thousands of time series. Fourth, the model better predictions. The ensemble always does better than the
presupposes that the training and testing datasets are competition, even when demand patterns change. It works
stationary. In real-world supply chain networks, though, best when the market is not very volatile (68.72%),
changes in the regime and shocks in demand can make moderately volatile (100.45%), or very volatile (49.52%).
stationarity not hold. Finally, demand time series with a lot The proposed framework is computationally efficient, taking
of zeros (more than 90%) make modeling harder, which only 23.8 minutes to train and 4.1 ms per series to infer.
shows how important it is to have models that can handle Because of this, it is good for big companies. Ablation
sparse and intermittent data. studies also show how important each part is. For instance,
. PatchTST is a good place to start for predictive performance,
residual learning helps you learn from your mistakes, and
There are many promising ways to improve and build on
adaptive weighting helps you get the best results for certain
the proposed framework. To begin with, we can apply the
series.
method to hierarchical forecasting, which needs to be
From a business perspective, the 28.5% MAPE reduction
consistent across different levels of product and geographic
enables substantial operational benefits: reduced inventory
locations. This is an important step in making the supply
carrying costs, improved service levels, and enhanced
chain work better in real life. Second, adding causal
planning efficiency. For typical mid-sized retailers, these
modeling and inference would let the framework show how
improvements translate to millions of dollars in annual cost
promotions, pricing plans, and other changes affect
savings.
outcomes, making the forecasts easier to understand and use.
Thirdly, creating probabilistic versions of PatchTST would Future research directions include extending to
give us full predictive distributions instead of point multivariate forecasting, developing probabilistic variants,
forecasts. This would give us better information about leveraging metalearning, and integrating causal inference.
uncertainty for tasks that come after. Fourth, meta-learning This work establishes error-adaptive ensembling as a
on several time series would help learn good ways to start, promising paradigm for next-generation demand forecasting
which would make it easier to quickly adapt to new or cold- systems, demonstrating that carefully designed hybrid
start products with little data. Fifth, using neural architecture architectures can overcome the limitations of both classical
search and other AutoML methods would make it possible to statistical methods and singlemodel deep learning
automatically find the best model architectures and approaches.
hyperparameters. Finally, adding ideas from robust REFERENCES
optimization would let the framework model forecast
[1] T. M. Choi, S. W. Wallace, and Y. Wang, “Big data analytics in
uncertainty directly, which would make supply chain operations management,” Production and Operations Management,
optimization more robust and aware of risk. vol. 27, no. 10, pp. 1868–1883, 2020.
[2] D. Ivanov and A. Dolgui, “Viability of intertwined supply networks:
D. Comparison with State-of-the-Art Extending the supply chain resilience angles towards survivability,”
International Journal of Production Research, vol. 58, no. 10, pp.
Our work contributes uniquely by: 2904– 2915, 2020.
• Focusing on highly volatile supply chain data [3] A. A. Syntetos, Z. Babai, J. E. Boylan, S. Kolassa, and K.
• Demonstrating ensemble benefits for intermittent Nikolopoulos, “Supply chain forecasting: Theory, practice, their gap
demand • Providing production-ready implementation and the future,” European Journal of Operational Research, vol. 252,
no. 1, pp. 1–26, 2016.
Our MAPE values (67-100%) appear higher than typical [4] T. Januschowski, J. Gasthaus, Y. Wang, D. Salinas, V. Flunkert, M.
benchmarks (∼10-20%), but this reflects the inherent Bohlke-Schneider, and L. Callot, “Criteria for classifying forecasting
methods,” International Journal of Forecasting, vol. 36, no. 1, pp. 167–
difficulty of supply chain forecasting with extreme volatility 177, 2020.
rather than model deficiency. Relative improvements over [5] G. E. Box and G. M. Jenkins, Time series analysis: Forecasting and
baselines (28.5%) align with gains reported in other control. Holden-Day, 1970.
ensemble forecasting studies [27]. [6] C. C. Holt, “Forecasting seasonals and trends by exponentially
weighted moving averages,” International Journal of Forecasting, vol.
VII. CONCLUSION 20, no. 1, pp. 5–10, 2004.
[7] S. Makridakis, E. Spiliotis, and V. Assimakopoulos, “The M4
This paper presents a novel error-adaptive ensemble competition: Results, findings, conclusion and way forward,”
framework for demand forecasting in supply chain International Journal of Forecasting, vol. 34, no. 4, pp. 802–808, 2018.
[8] S. J. Taylor and B. Letham, “Forecasting at scale,” The American [29] A. Timmermann, “Forecast combinations,” in Handbook of Economic
Statistician, vol. 72, no. 1, pp. 37–45, 2018. Forecasting, vol. 1. Elsevier, 2006, pp. 135–196.
[9] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural [30] J. M. Bates and C. W. J. Granger, “The combination of forecasts,”
Computation, vol. 9, no. 8, pp. 1735–1780, 1997. Journal of the Operational Research Society, vol. 20, no. 4, pp. 451–
[10] H. Hewamalage, C. Bergmeir, and K. Bandara, “Recurrent neural 468, 1969.
networks for time series forecasting: Current status and future [31] P. Newbold and C. W. J. Granger, “Experience with forecasting
directions,” International Journal of Forecasting, vol. 37, no. 1, pp. univariate time series and the combination of forecasts,” Journal of the
388–427, 2021. Royal Statistical Society: Series A (General), vol. 137, no. 2, pp. 131–
[11] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. 146, 1974.
Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in [32] D. H. Wolpert, “Stacked generalization,” Neural Networks, vol. 5, no.
Advances in Neural Information Processing Systems, 2017, pp. 5998– 2, pp. 241–259, 1992.
6008. [33] J. H. Stock and M. W. Watson, “Combination forecasts of output
[12] Y. Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series growth in a seven-country data set,” Journal of Forecasting, vol. 23,
is worth 64 words: Long-term forecasting with transformers,” in no. 6, pp. 405–430, 2004.
International Conference on Learning Representations, 2023. [34] J. H. Friedman, “Greedy function approximation: A gradient boosting
[13] F. Petropoulos et al., “Forecasting: Theory and practice,” International machine,” Annals of Statistics, vol. 29, no. 5, pp. 1189–1232, 2001.
Journal of Forecasting, vol. 38, no. 3, pp. 705–871, 2022. [35] T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,”
[14] S. Smyl, “A hybrid method of exponential smoothing and recurrent in Proceedings of the 22nd ACM SIGKDD International Conference
neural networks for time series forecasting,” International Journal of on Knowledge Discovery and Data Mining, 2016, pp. 785–794.
Forecasting, vol. 36, no. 1, pp. 75–85, 2020. [36] G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T. Y.
[15] S. Kolassa, “Evaluating predictive count data distributions in retail Liu, “LightGBM: A highly efficient gradient boosting decision tree,”
sales forecasting,” International Journal of Forecasting, vol. 32, no. 3, in Advances in Neural Information Processing Systems, 2017, pp.
pp. 788–803, 2016. 3146–3154.
[16] P. Montero-Manso and R. J. Hyndman, “Principles and algorithms for [37] V. Cerqueira, L. Torgo, and I. Mozetic, “Evaluating time series
forecasting groups of time series: Locality and globality,” International forecast-ˇ ing models: An empirical study on performance estimation
Journal of Forecasting, vol. 37, no. 4, pp. 1632–1653, 2021. methods,” Machine Learning, vol. 109, no. 11, pp. 1997–2028, 2020.
[17] R. J. Hyndman and G. Athanasopoulos, Forecasting: Principles and [38] D. P. Kingma and J. Ba, “Adam: A method for stochastic
practice, 2nd ed. OTexts, 2018. optimization,” in International Conference on Learning
[18] Z. Ma and T. Pan, "Adaptive Weight Tuning of EWMA Controller via Representations, 2015.
Model-Free Deep Reinforcement Learning," IEEE Transactions on
Semiconductor Manufacturing, vol. 36, no. 1, pp. 1–9, Feb. 2023, doi:
10.1109/TSM.2023.3236798.
[19] L. Kumar, S. Khedlekar, and U. K. Khedlekar, "A comparative
assessment of Holt-Winters exponential smoothing and autoregressive
integrated moving average for inventory optimization in supply
chains," Supply Chain Analytics, vol. 4, art. no. 100084, Sep. 2024,
doi: 10.1016/[Link].2024.100084.
[20] S. Makridakis, E. Spiliotis, and V. Assimakopoulos, “The M4
competition: 100,000 time series and 61 forecasting methods,”
International Journal of Forecasting, vol. 36, no. 1, pp. 54–74, 2020.
[21] K. Cho, B. Van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares,¨
H. Schwenk, and Y. Bengio, “Learning phrase representations using
RNN encoder-decoder for statistical machine translation,” arXiv
preprint arXiv:1406.1078, 2014.
[22] D. Salinas, V. Flunkert, J. Gasthaus, and T. Januschowski, “DeepAR:
Probabilistic forecasting with autoregressive recurrent networks,”
International Journal of Forecasting, vol. 36, no. 3, pp. 1181–1191,
2020.
[23] S. Bai, J. Z. Kolter, and V. Koltun, “An empirical evaluation of generic
convolutional and recurrent networks for sequence modeling,” arXiv
preprint arXiv:1803.01271, 2018.
[24] A. van den Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals, A.
Graves, N. Kalchbrenner, A. Senior, and K. Kavukcuoglu, “WaveNet:
A generative model for raw audio,” arXiv preprint arXiv:1609.03499,
2016.
[25] B. Lim, S. O. Arık, N. Loeff, and T. Pfister, “Temporal fusion
transform-¨ ers for interpretable multi-horizon time series forecasting,”
International Journal of Forecasting, vol. 37, no. 4, pp. 1748–1764,
2021.
[26] H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and [Link],
Informer: Beyond efficient transformer for long sequence timeseries
forecasting,” in Proceedings of the AAAI Conference on Artificial
Intelligence, vol. 35, no. 12, 2021, pp. 11106–11115.
[27] T. G. Dietterich, “Ensemble methods in machine learning,” in
International Workshop on Multiple Classifier Systems. Springer,
2000, pp. 1–15.
[28] R. T. Clemen, “Combining forecasts: A review and annotated
bibliography,” International Journal of Forecasting, vol. 5, no. 4, pp.
559–583, 1989.