Structured Data AI Model Selection Guide: Tabular & Time Series for Industrial Operations

Overview

This guide shows you how to go from operational problem to deployed model with a simple path:

1. Identify your problem
2. Check your data situation
3. Start with a proven baseline, then evaluate modern alternatives

It covers the common industrial workloads — demand forecasting, equipment failure prediction, defect classification, energy optimization, and risk scoring — and offers options for data-rich, data-scarce, and zero-data scenarios.

The best model is worthless with poor features or a mis-specified problem. This guide assumes you've done the data work. It makes the model selection step systematic so you can move faster and argue less.

Once you have chosen a model, use the Readiness Checklist further below to verify that the model fits your operational constraints.

Before you commit budget → Jump to the Readiness Checklist to confirm your candidate model fits your problem, data, infrastructure, and team. It's the single fastest way to catch mismatches before they become expensive.

Table of Contents

1. Overview
2. Know Your Data Before You Pick Your Model
   • Tabular Data Types
   • Why This Matters for Model Selection
   • Data Frequency Also Matters
3. Decision Tree: From Problem to Model
   • Step 1 — What type of data do you have?
   • Step 2A — Tabular Data: How much do you have?
   • Step 2B — Time Series: Do you have training data?
4. Mapping Models to Operational Problems
5. Deployment Examples
6. Constraints and Cost Traps
7. Implementation Roadmap
8. Readiness Checklist
9. What's New (and What Isn't)
10. References

Know Your Data Before You Pick Your Model

Before choosing a model, identify what type of structured data you have. This determines which models apply and which questions the data can actually answer.

Tabular Data Types

Type	What It Is	Industrial Example	What It Can Answer
Cross-sectional	Many subjects observed at one point in time. Each row is a different unit (machine, customer, plant).	A snapshot of all machines in a plant with their age, operating hours, and defect count — taken today.	Questions about levels and differences: "Which machines are highest-risk right now?"
Repeated cross-section	The same survey or measurement administered to different samples at successive time points.	Annual supplier quality audits where different suppliers are sampled each year.	Questions about trends: "Is supplier quality improving or declining across the portfolio?"
Time series	One subject measured at multiple points in time, typically at regular intervals (hourly, daily, monthly).	Hourly electricity consumption at a single plant over two years.	Questions about patterns and forecasting: "Is there a seasonal component in our energy costs?"
Panel data	The same subjects observed over time. Each row is a subject-time combination (e.g., machine-month).	Monthly sensor readings for every turbine in your fleet, tracked over three years.	Questions about change and causality: "Which turbines are degrading faster, and why?"

Why This Matters for Model Selection

• Cross-sectional data → Tabular models. Start with XGBoost/LightGBM/CatBoost ¹, ²; evaluate TabPFN-2.5 or TabICLv2 as alternatives, especially on small datasets. One row per unit, predict a label or score.
• Time series data → Time series models (NHITS, TimesFM, Chronos). Sequence matters; the model learns temporal patterns.
• Panel data → Either approach, depending on the question. Predict per-unit outcomes with tabular models, or forecast per-unit trajectories with time series models.
• Repeated cross-sections → Tabular models with temporal drift handling (Drift-Resilient TabPFN ³) if the distribution shifts between measurement periods.

If you're unsure which type you have, ask: "Are my rows different units at one time, or the same unit at different times?" That single question determines your path through the decision tree below.

Data Frequency Also Matters

For time series data, sampling frequency narrows the field further:

Frequency	Examples	Best Fit
High (< 1 minute)	Vibration sensors, tick data, IoT streams	Neural models: NHITS 4, PatchTST 5
Medium (hourly–daily)	Energy meters, production counts, weather	Foundation or neural: TimesFM 6, NHITS 4
Low (weekly–monthly)	Sales, financial reporting, inspections	Foundation or statistical: TimeGPT 7, Prophet 8
Irregular (event-driven)	Maintenance logs, fault events	Chronos 9 (handles irregular sampling)
Multiple correlated series	Multi-sensor arrays, fleet-wide data	MOMENT 10, TimeGPT 7

Decision Tree: From Problem to Model

Data first, then model. No model compensates for misspecified problems, poor features, or dirty data. Before entering this tree, confirm that you have a clearly defined prediction target, that your data actually measures what you think it measures, and that someone on the team understands the operational context well enough to translate model outputs into decisions.

Step 1 — What type of data do you have?

Data Type	Description	Go To
Tables	Rows and columns — ERP exports, inspection logs, customer records, financial data	Step 2A
Time series	Temporal sequences — sensor streams, demand history, energy consumption, price data	Step 2B

Step 2A — Tabular Data: How much do you have?

Start with a gradient boosting baseline. For any tabular dataset above ~1,000 rows, XGBoost, LightGBM, or CatBoost ¹, ² should be your first experiment. They are fast to train on CPU, handle mixed data and missing values natively, and remain the dominant approach in production and competitive benchmarks. The foundation models below are valuable alternatives and complements — not replacements.

Dataset Size	Robust Baseline	Advanced Alternative	Time to First Result*
Small (< 10K rows)	XGBoost / LightGBM 1	TabPFN-2.5 11 or TabICLv2 12 (zero-shot, often competitive without tuning)	Days
Medium (10K–50K rows)	XGBoost / LightGBM 1	TabICLv2 12 or TabPFN-2.5 11	Days
Large (50K–500K rows)	XGBoost / LightGBM 1	TabICLv2 12 or Chunked-TabPFN 13	Days–weeks
Very large (500K–10M rows)	XGBoost / LightGBM 1	Chunked-TabPFN 13	Weeks
Massive (> 10M rows)	XGBoost / LightGBM 1	—	Weeks
Mixed numeric + text	CatBoost 2 or embeddings + XGBoost	FT-TabPFN 14	Days
High-cardinality categoricals	CatBoost 2	—	Days–weeks

*"Time to First Result" refers to the full project cycle (data cleaning, validation, deployment) — not model inference. Foundation models like TabPFN return predictions in seconds to minutes; the surrounding work takes longer.

Where foundation models shine: TabPFN-2.5 achieves a 100% win rate against default (untuned) XGBoost on classification datasets up to 10,000 rows and 500 features, and an 87% win rate on larger datasets up to 100,000 rows — with zero hyperparameter tuning ¹¹. Their advantage is strongest when you need a fast, defensible result without a tuning cycle.

Where gradient boosting holds: With proper hyperparameter tuning, XGBoost and LightGBM close much of that gap and often win on medium-to-large datasets ¹. In most Kaggle competitions and open ML benchmarks, tuned gradient boosting remains the dominant method for standard supervised problems.

Important caveat: Both TabPFN and TabICLv2 benchmarks were run under specific conditions. TabPFN's headline results compare against untuned XGBoost baselines ¹¹. TabICLv2's claims (February 2026) are from the authors' own benchmarks and have not yet been independently reproduced; the comparison baseline used TabPFN-2.5 with additional tuning and ensembling ¹². Evaluate both against a properly tuned gradient boosting baseline on your data.

Note: TabICLv2 also supports zero-shot time series forecasting via TabICLForecaster ¹². If you adopt it for tabular work, you get a forecasting option from the same tool without adding a second dependency.

Step 2B — Time Series: Do you have training data?

Data Situation	Priority	Recommended Model	Key Advantage
No training data	Speed	TimesFM 6	Up to 179× faster than similarly-sized Chronos on benchmarked tasks; near-SOTA zero-shot 6, 15
No training data	Uncertainty estimates	Chronos 9	19–60% CRPS reduction on load forecasting 16
No training data	No infrastructure	TimeGPT 7	API-based, no GPU required 7
No training data	Long multivariate sensor data	MOMENT 10	Compressive memory for extended cross-channel context 10
Training data available	Long horizon + speed	NHITS 4	~20% accuracy gain, ~50× speedup vs Transformers 4
Training data available	Interpretability	N-BEATS 17	Explicit trend/seasonality decomposition 17
Training data available	Long look-back	PatchTST 5	21% MSE reduction, 22× faster on large datasets 5
Training data available	Multiple input variables	TFT 18	Variable importance scoring built in 18
Training data available	Simple baseline	Prophet 8	Fast, interpretable, low compute 8

Mapping Models to Operational Problems

Choose your problem. Apply your constraint. Select from the table.

Operational Problem	Data Available?	Baseline Approach	Advanced Alternative	Needs Dedicated ML Team?
Equipment failure prediction	Yes (sensor/inspection logs)	XGBoost on engineered features	NHITS 4 or PatchTST 5	Low–Moderate
Remaining useful life (RUL) estimation	Yes (run-to-failure history)	Survival analysis or XGBoost	NHITS 4 with multi-horizon output	Moderate
Anomaly detection in sensor streams	Yes (normal operation data)	Statistical process control	MOMENT 10 or Chronos 9	Moderate
Demand forecasting (existing line)	Yes (ERP history)	Prophet 8 or ARIMA	NHITS 4	Low–Moderate
Demand forecasting (new business / new domain)	No	—	TimesFM 6 or TimeGPT 7	Low
Defect classification	Limited (few examples)	XGBoost / LightGBM 1	TabPFN-2.5 11 or TabICLv2 12	Low
Quality scoring (continuous)	Yes (inspection records)	XGBoost / LightGBM 1	TabICLv2 12 or TabPFN-2.5 11	Low
Cost / risk scoring	Yes (structured tables)	XGBoost / LightGBM 1	TabICLv2 12 or TabPFN-2.5 11	Low
Energy consumption optimization	Yes (meter/sensor data)	Prophet 8	N-BEATS + TFT 19	Moderate
Long-horizon resource planning	Yes (historical series)	ARIMA / Prophet 8	PatchTST 5	Moderate
Multi-sensor monitoring (vibration, temp, pressure)	Yes (multi-channel streams)	Statistical process control	MOMENT 10	Moderate
Classification with text fields	Yes (mixed tables)	Embeddings + XGBoost, or CatBoost 2	FT-TabPFN 14	Low–Moderate
Quality control (new product line)	Limited	XGBoost 1	TabPFN-2.5 11	Low

Deployment Examples

1. Equipment Failure Prediction — Railway Operations

Hitachi deployed TabPFN to predict component failures in its rail network ²⁰. The problem: specific failure modes (e.g., brake pad wear, signal relay faults) occur infrequently — sometimes only 10–20 times per year across thousands of components. Traditional models struggle with this class imbalance. TabPFN excels on small-data scenarios where a specific failure mode has limited historical examples ²¹. The outcome: reduced unplanned downtime by identifying at-risk components before failure, without waiting years to accumulate training data.

2. Energy Forecasting — Interpretable for Stakeholders

A traction energy forecasting study combined N-BEATS with Temporal Fusion Transformers, achieving RMSE of 0.06 with quantified external factor importance ¹⁹. N-BEATS shows why the forecast says what it says ¹⁷. TFT identifies which external factors drive consumption ¹⁸. A forecasting model that your operations team actually trusts — because they can see the decomposition — gets adopted. A black box gets ignored.

3. Demand Forecasting — No Unified Data

Foundation models address a common integration problem: fragmented legacy systems, no unified history, and a planning cycle that won't wait.

TimeGPT demonstrated competitive zero-shot accuracy on soil moisture forecasting using only historical measurements ²². TimesFM was fine-tuned on 100 million financial time points to improve price prediction accuracy ²³. Both illustrate the same principle: pretrained models give you a defensible starting point without waiting months for data cleanup.

Constraints and Cost Traps

Verify these constraints before committing budget.

Constraint	What to Watch	Source
Real-time latency required	Do not deploy Chronos or Lag-Llama — both are >600× slower than LSTM baselines. Use TimesFM (up to 179× faster than similarly-sized Chronos) or NHITS.	15, 4
Very large datasets (>10M rows)	XGBoost/LightGBM still win on scalability and cost. Don't pay GPU costs for a problem commodity hardware solves.	1
Missing data	TabPFN requires complete data — missing values must be imputed before inference. High-cardinality categoricals require preprocessing.	21
Unverified vendor claims	TabICLv2's SOTA claims have not yet been independently reproduced. The comparison baseline used TabPFN-2.5 with additional tuning and ensembling.	12
No baseline established	Don't skip Phase 1 (assessment) and Phase 2 (baseline). If someone proposes jumping straight to foundation or neural models without establishing what Prophet/ARIMA (time series) or tuned XGBoost (tabular) can do first, they're selling you hours, not outcomes.	1, 8

Implementation Roadmap

Don't skip phases. Each one takes roughly a week.

Phase	What You Do	Why It Matters
1. Assessment	Characterize data (type, size, frequency, quality). Define accuracy, speed, and interpretability requirements.	Prevents selecting a model that can't run on your data or infrastructure.
2. Baseline	Implement Prophet or ARIMA for time series 8; XGBoost or LightGBM for tabular 1. Establish performance metrics.	Gives you a number to beat. If someone proposes skipping this, push back.
3. Foundation models	Try zero-shot with TimesFM, Chronos, or TimeGPT (time series) 6, 9, 7, or TabPFN-2.5 / TabICLv2 (tabular) 11, 12.	Fastest way to see what's achievable without training.
4. Neural models	Train NHITS, PatchTST, or TFT if sufficient data exists 4, 5, 18. Compare to Phase 2 and 3.	Often the accuracy ceiling — but only if data quality and volume justify it.
5. Production	Select best model. Build monitoring and retraining pipeline. Deploy.	A model without drift monitoring and a retraining schedule is a liability, not an asset.

Combine models when it makes sense. Ensembles often outperform single models. The source guide documents N-BEATS + TFT achieving RMSE of 0.06 in energy forecasting ¹⁹ — better than either model alone. A common pattern: use a foundation model for the initial estimate, then fine-tune or ensemble with a trained neural model as data accumulates.

Readiness Checklist

Use this checklist to confirm a candidate model fits your problem and environment before committing budget.

Property	Description	✓
Problem match	Model supports the required task (forecasting, classification, scoring)	☐
Data readiness	Data is clean, complete, and accessible — or a zero-shot model is selected	☐
Accuracy	Model reaches required accuracy on your validation data or published benchmarks	☐
Latency	Model runs fast enough for your operational cadence (real-time vs. batch)	☐
Hardware fit	Model fits into memory of target hardware (GPU, CPU, edge)	☐
Interpretability	Outputs are explainable to the stakeholders who must act on them	☐
Baseline comparison	Performance has been compared against a simple baseline (Prophet, XGBoost)	☐
Maintenance plan	Retraining cadence defined (foundation models: none; neural models: monthly/quarterly)	☐
Drift monitoring	Plan exists to detect when model performance degrades over time	☐
License	Code and weights license permits commercial use	☐
Team capability	Team can deploy and maintain, or a qualified partner is identified	☐

What's New (and What Isn't)

Foundation models have dramatically reduced the biggest bottleneck in industrial AI: the months of dataset-specific hyperparameter tuning that used to make every project a gamble ²¹, ⁶. The bottleneck hasn't disappeared — it has shifted from hyperparameter search to data preparation, prompt design, and inference configuration — but the barrier to a first defensible result is far lower.

Four things are different now:

1. Forecasting and classification are deployable in weeks, not quarters, if you have clean historical data ⁴, ⁶.

2. Data-scarce scenarios no longer require waiting for data collection — zero-shot models provide defensible first estimates immediately ⁶, ⁷.

3. Small-data problems (rare defects, limited labeled examples, new product lines) that were previously unsolvable without massive datasets are now tractable ²¹, ²⁴.

4. The cost structure of experimentation has changed. Foundation models are pretrained — you pay only for inference, not training ²¹, ⁶, ⁹. But inference costs for large models (especially on GPU) can exceed training costs for simpler methods. Evaluate total cost, not just model training cost.

Two things haven't changed: you still need someone who understands the problem, can assess data quality, and can translate model outputs into decisions. And gradient boosting on well-engineered features remains the most reliable default for standard supervised tabular problems ¹. The new models expand what's possible. They don't obsolete what already works.

References

Footnotes

1. T. Chen and C. Guestrin, "XGBoost: A scalable tree boosting system," in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 785-794, 2016. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴ ↩¹⁵ ↩¹⁶

2. L. Prokhorenkova, G. Gusev, A. Vorobev, A. V. Dorogush, and A. Gulin, "CatBoost: unbiased boosting with categorical features," in Advances in Neural Information Processing Systems, vol. 31, 2018. ↩ ↩² ↩³ ↩⁴ ↩⁵

3. B. Helli, S. Müller, N. Hollmann, and F. Hutter, "Drift-Resilient TabPFN: In-Context Learning Temporal Distribution Shifts on Tabular Data," arXiv:2411.10634, 2024. ↩

4. C. Challu, K. G. Olivares, B. N. Oreshkin, F. Garza, M. Mergenthaler-Canseco, and A. Dubrawski, "NHITS: Neural Hierarchical Interpolation for Time Series Forecasting," Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 6, pp. 6989-6997, 2023. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰

5. Y. Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, "A Time Series is Worth 64 Words: Long-term Forecasting with Transformers," International Conference on Learning Representations, 2022. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

6. A. Das, W. Kong, A. Leach, S. Mathur, R. Sen, and Y. Yu, "A decoder-only foundation model for time-series forecasting," arXiv:2310.10688, 2023. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

7. A. Garza and M. Mergenthaler-Canseco, "TimeGPT-1," arXiv:2310.03589, 2023. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

8. S. J. Taylor and B. Letham, "Forecasting at scale," The American Statistician, vol. 72, no. 1, pp. 37-45, 2018. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

9. A. Ansari, L. Stella, C. Turkmen, X. Zhang, et al., "Chronos: Learning the Language of Time Series," arXiv:2403.07815, 2024. ↩ ↩² ↩³ ↩⁴ ↩⁵

10. M. Zukowska, O. Melnyk, M. Moor, and T. Palpanas, "Towards Long-Context Time Series Foundation Models," arXiv:2409.13530, 2024. ↩ ↩² ↩³ ↩⁴ ↩⁵

11. N. Hollmann, S. Müller, and F. Hutter, "TabPFN: Accurate Predictions on Small Data with a Tabular Foundation Model," arXiv:2511.08667, November 2025. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

12. J. Qu, D. Holzmüller, G. Varoquaux, and M. Le Morvan, "TabICLv2: A better, faster, scalable, and open tabular foundation model," arXiv:2602.11139, February 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰

13. R. Sergazinov, A. Shen, S. Müller, F. Hutter, and A. Dubrawski, "Chunked TabPFN: Exact Training-Free In-Context Learning for Long-Context Tabular Data," 2025. ↩ ↩²

14. Y. Liu, S. Müller, and F. Hutter, "Tokenize features, enhancing tables: the FT-TABPFN model for tabular classification," arXiv:2406.06891, 2024. ↩ ↩²

15. S. Ali, A. Alvi, S. Raza, and M. Yousuf, "Zero-shot forecasting for ECG time series data using generative foundation models," in 2024 IEEE International Conference on Body Sensor Networks (BSN), pp. 1-4, 2024. ↩ ↩²

16. Z. Liao, K. Liang, K. Xu, and B. Cui, "Zero-Shot Load Forecasting with Large Language Models," arXiv:2411.11350, 2024. ↩

17. B. N. Oreshkin, D. Carpov, N. Chapados, and Y. Bengio, "N-BEATS: Neural basis expansion analysis for interpretable time series forecasting," in International Conference on Learning Representations, 2020. ↩ ↩² ↩³

18. B. Lim, S. Ö. Arık, N. Loeff, and T. Pfister, "Temporal Fusion Transformers for interpretable multi-horizon time series forecasting," International Journal of Forecasting, vol. 37, no. 4, pp. 1748-1764, 2021. ↩ ↩² ↩³ ↩⁴

19. Y. Jiang, Y. Zhao, Y. Guo, and Y. Jiang, "Interpretable Forecasting of Traction Energy Consumption Based on Nbeats and Temporal Fusion Transformers," in 2024 IEEE 7th International Conference on Industrial Cyber-Physical Systems (ICPS), pp. 1-6, 2024. ↩ ↩² ↩³

20. "How Hitachi Uses TabPFN for Equipment Failure Prediction," Prior Labs Case Studies / Hitachi partnership announcement. ↩

21. N. Hollmann, S. Müller, K. Eggensperger, and F. Hutter, "Accurate predictions on small data with a tabular foundation model," Nature, vol. 635, pp. 115-121, January 2024. ↩ ↩² ↩³ ↩⁴ ↩⁵

22. L. Deforce, B. Masseran, T. Voisin, and A. Bozzon, "Leveraging Time-Series Foundation Models in Smart Agriculture for Soil Moisture Forecasting," arXiv:2405.18913, 2024. ↩

23. Y. Fu, Y. Xiong, Y. Tian, S. Zhang, et al., "Financial Fine-tuning a Large Time Series Model," arXiv:2412.09880, 2024. ↩

24. "How BostonGene Utilized TabPFN to Identify Immune System Profiles," Prior Labs Case Studies. ↩