Tell us which forecasting model you're most interested in and we'll notify you when it's ready.
We'll reach out to when your selected model is ready for early access.
From classical statistics to foundation models — a research platform unifying TimeGPT, Chronos, MOIRAI, and hybrid deep learning architectures for supply chain and retail demand intelligence.
Describe your forecasting challenge — industry, data availability, horizon, and constraints — and get an instant, research-backed recommendation.
A new paradigm: pre-trained foundation models that achieve zero-shot and few-shot forecasting across any domain — no retraining required.
The world's first foundation model for time series forecasting and anomaly detection. Trained on over 100 billion data points spanning finance, weather, retail, energy, and healthcare. Achieves state-of-the-art zero-shot performance — no fine-tuning required for new datasets.
Our in-house research model fusing classical statistical methods (ARIMA, ETS, Grey Models) with large language model reasoning. StatisticalGPT interprets demand signals in natural language, provides explainable forecasts, and automatically selects the optimal statistical model family per SKU.
Probabilistic forecasting via language model pre-training on a large corpus of real and synthetic time series. Quantizes time series into tokens and trains a T5 encoder-decoder. Outperforms specialized models on most zero-shot benchmarks.
Unified Training of Universal Time Series Forecasting Transformers. Trains on LOTSA — a dataset of 27 billion observations across 9 domains. Supports any-variate forecasting with patch-based tokenization and flexible distribution heads.
A 200M parameter decoder-only foundation model pre-trained on 100B real-world time-series data points. Uses patching and masked pre-training. Achieves near-supervised performance in zero-shot settings on M4, ETT, and Weather benchmarks.
First open-source foundation model for probabilistic time series forecasting built on LLaMA architecture. Uses lag-based tokenization. Achieves strong zero-shot performance and can be fine-tuned with minimal data — ideal for cold-start SCM scenarios.
NHITS uses hierarchical interpolation for long-horizon forecasting. PatchTST applies channel-independent patching to transformers. Both set new records on ETT and M4 benchmarks — central to the Notion Demand neural forecasting stack.
Decomposable multiscale mixing architecture — separately models seasonal and trend components at multiple resolutions. Outperforms iTransformer and PatchTST on several long-horizon benchmarks with significantly fewer parameters.
From classical statistical baselines to cutting-edge hybrid deep learning architectures — comprehensive analysis of demand forecasting approaches across supply chain and retail domains.
Multi-Channel Data Fusion Network
Hybrid CNN + BiLSTM + GRU architecture that extracts both spatial and temporal features simultaneously. Outperforms 7 other deep learning models with statistically verified results (paired t-test, p=0.05).
Dynamic LSTM Supply Chain Management
Dynamically updates deployed LSTM models using historical sales data augmented with lag and rolling window statistical features. Evaluated on Walmart benchmark dataset.
Hybrid Ensemble Stacking
Novel stacking ensemble combining Random Forest bagging + XGBoost boosting with Logistic Regression meta-learner. Achieves R² = 0.9651 on real-time retail sales data.
Fourier Time-Varying Grey Model
Grey model extended with Fourier functions for seasonal pattern capture. Data-driven order selection algorithm. Outperforms SARIMA, LSTM, MLP, and Holt-Winters on M5 competition data.
Multi-Layer Perceptron / Artificial Neural Network
Most widely used architecture in demand forecasting literature (2017–2021). Forms a key baseline across comparative studies in supply chain and retail contexts.
ARIMA · SARIMA · Holt-Winters · ETS
Classical benchmarks against which ML models are evaluated. SARIMA and Holt-Winters serve as standard seasonal baselines in all time-series competition frameworks.
Peer-reviewed publications spanning supply chain optimization, retail forecasting, big data analytics, and cutting-edge hybrid deep learning architectures.
Introduces MCDFN, a hybrid deep learning architecture that integrates CNN, LSTM, and GRU to extract spatial and temporal features. Verified by paired t-test (p=0.05) and 10-fold cross-validation. Enhanced with ShapTime and Permutation Feature Importance for explainability.
Proposes a framework that dynamically updates LSTM models using historical sales data and lag/rolling-window features. Evaluated on the Walmart M5 dataset. Multi-layer LSTM with lag features achieves significantly lower RMSE than single-layer baselines.
Read Paper →Comprehensive comparison of RF, ANN, GB, AdaBoost, XGBoost vs. the novel hybrid RF-XGBoost-LR. Stacking meta-learner combines bagging and boosting strengths. Best performance: R² = 0.9651, MAE = 0.0025.
Proposes FTGM extending grey models with Fourier functions for seasonal variation capture. Data-driven order selection. Outperforms SARIMA, HWES, MLP, LSTM, and DGSM on M5 competition data (70 monthly product demand time series, 7 departments, 10 stores).
Read Paper →Analyzes 23 AI methods published 2017–2021 using Web of Science, IEEE Explore, and Springer. Classifies by dimensionality, data volume, and forecast horizon. Identifies trend toward deep learning (MLP, LSTM, ANN) and gap in "collaborative forecasting".
Read Paper →Surveys predictive BDA in SCM demand forecasting (2005–2019). Classifies into 7 technique families. Identifies neural networks and regression as most common. Highlights major gap: no BDA research on closed-loop supply chains (CLSCs).
Read Paper →Investigates calendric effects (holidays, promotions, special events) on retail demand forecasting accuracy. Demonstrates that incorporating event-calendar features into ML models significantly improves daily-granularity predictions.
Read Paper →Explores combinatorial deep learning architectures for SCM demand forecasting. Evaluates CNN, LSTM, GRU combinations in a systematic fashion, identifying optimal fusion strategies for multi-horizon supply chain predictions.
Read Paper →Compiled from peer-reviewed results across supply chain and retail demand forecasting datasets — traditional and foundation models side by side.
| Model | Type | RMSE ↓ | MAE ↓ | MAPE ↓ | R² ↑ | Dataset |
|---|---|---|---|---|---|---|
| TimeGPT-1 Foundation | Transformer (Zero-Shot) | SOTA | SOTA | SOTA | — | Multiple (100B pts) |
| MCDFN Best Hybrid | CNN+LSTM+GRU | 4.86% | 3.99% | 20.16% | — | SC Retail (BD) |
| Chronos | T5 Transformer (Zero-Shot) | Best WQL | Best WQL | — | — | Multiple (0-shot) |
| RF-XGBoost-LR | Hybrid Ensemble | — | 0.0025 | — | 0.9651 | Walmart US |
| FTGM | Grey + Fourier | Rank #1 | Rank #1 | — | — | M5 Competition |
| DLSTM-SCM | Dynamic LSTM | Improved | Improved | — | — | Walmart M5 |
| Vanilla LSTM | RNN | Higher | Higher | Higher | — | Multiple |
| SARIMA | Statistical | 26.6% worse than CNN | — | — | — | M5 Competition |
Join researchers and supply chain teams who are building with the next generation of AI forecasting models. Select your model of interest and we'll keep you in the loop.
No spam. Unsubscribe anytime.
Hybrid architectures (CNN+LSTM+GRU, RF+XGBoost+LR) consistently outperform single-architecture models across all studied datasets and error metrics.
Adding lag features (1-day, 7-day, 28-day) and rolling window statistics to LSTM models reduces both training and validation loss significantly across all SCM datasets.
Incorporating holiday calendars, promotional events, and special-day indicators into ML models significantly improves daily retail demand forecasting accuracy.
Adding exogenous variables (weather, macroeconomic indicators, location data) to deep learning models reduces RMSE by up to 50.9%, validating multivariate over univariate approaches.
Literature almost entirely focuses on retailer-perspective forecasting. Upstream "collaborative forecasting" involving suppliers and manufacturers remains a critical research gap.
TimeGPT, Chronos, and MOIRAI are redefining the benchmark — achieving near-supervised accuracy without any task-specific training, lowering the barrier to AI forecasting for SMEs.