The One-Degree Dispatch

Insulation Demand Forecasting POV Report v1.7 07-01-25 - Read-Only

2025 · Causal AI · 3,324 words

This report unveils how Causal AI optimizes insulation demand forecasting, transforming Owens Corning’s supply chain for smarter enterprise operations.

Building Smarter Enterprises with Causal AI 1

Parabole | Owens Corning: Insulation Demand Forecasting POV Report

1st July 2025

POV Team

Owens Corning

Parabole

Paul Bouvy – Senior Director of Supply Chain

Rajib Saha – Chief Executive Officer

Margaret Hoff – Supply Chain CI Leader

Sandip Bhaumik – Chief Technical Officer

Michael Gibbs – Material Planning Manager

Shubho Ghosh – Advisor and Execution Lead Sai Anvesh Durvasula – Senior Architect Nilanjan Gantait – Senior Architect

Sree Charan Pothireddi – Lead Data Scientist Sagiruddin Mondal – Lead Engineer

Executive summary

Objective & background of the problem

This POV was initiated to assess how Causal AI can enhance Owens Corning’s Insulation S&OP team’s ability to forecast demand more accurately by explicitly modeling the cause-and-effect relationships that drive insulation material demand. The team currently relies on statistical forecasting methods rooted in historical patterns, which limits their ability to assess the impact of strategic decisions or adapt to novel market conditions. By leveraging Parabole’s TRAIN platform, the goal was to move beyond correlation-based approaches and develop an explanatory causal model that reflects real-world drivers of demand. This would enable the team to simulate counterfactual scenarios—such as regulatory changes, shifts in construction activity, or macroeconomic movements—and use those insights to support strategic planning and improve forecast reliability under dynamic conditions.

Accomplished solution and benefits

The TRAIN platform generated a causal model that linked external and internal demand drivers —such as housing starts, energy prices, regulatory changes, and consumer behavior trends—to market-level demand for insulation materials. Unlike traditional statistical models, it captured complex interdependencies betwe en these factors and uncovered the underlying cause-and-effect logic driving demand fluctuations. By integrating SME knowledge with structured and unstructured data sources, the platform d elivered improved forecast accuracy even in conditions of data sparsity and market volatility. As part of the solution, a Demand Forecasting Agent was developed to simulate hypothetical scenarios—such as changes in mortgage rates, extreme weather events, or construction activity changes—and assess their projected impact on overall market demand. This enables the S&OP team to shift from reactive forecasting to forward-looking, scenario-based planning. The causal model was modular, scalable to adjacent categories or regions, and offered transparent, explainable outputs that fostered cro ss-functional alignment and more confident decision-making.

Execution summary

The POV was completed in 15 weeks, with under 60 hours of machine effort for model creation and minimal SME involvement. The platform supported incremental knowledge model building, iterative updates to data without retraining and produced an agent capable of simulating real-world demand scenarios.

Key learnings

Causal AI proved effective in bridging data gaps by uncovering relationships that traditional models often overlook. The inte gration of SME insights helped surface weak signals and domain-specific nuance, while the explainable nature of the model improved trust across teams. Critically, the project showed that accurate f orecasting can be achieved using a small set of truly causal factors— reducing reliance on large, complex datasets and enabling more focused, interpretable, and resilient models.

Challenges faced

Data limitations in regulatory, energy, socio-political, and warehousing/supply chain constrained certain parts of the model, requiring model to be trained on available input.

Potential next steps

A next step could involve extending the causal model to SKU-level demand forecasting by incorporating supply-side, inventory, and sales team inputs—enabling more granular and actionable planning. In parallel, the platform can be applied to manufacturing use cases to discover causal factors behind material qual ity deviations. This would support a shift toward proactive, explainable root-cause diagnostics. Together, these applications demonstrate the broader potential of causal AI across planning and operatio ns.

Objectives, Scope, and Context

Aim: To prove how Causal AI can enhance Owens Corning’s Insulation S&OP team’s ability to model cause-and-effect relationships, enabling more accurate demand forecasting and better-informed strategic decision-making

Background

Owens Corning’s Insulation S&OP team primarily relies on statistical forecasting techniques that analyze historical data patterns, such as trends and seasonality, to predict demand for insulation materials. While effective for recognizing correlations, these models struggle to adapt to unprecedented events or assess the impact of strategic changes. Without an explicit understanding of causal relationships, the team faces challenges in evaluating the effects of key business decisions beyond historical precedent. Recognizing these limitations, the team is interested in use Causal AI analytics to determine its potential for improving demand forecasting. By moving beyond traditional pattern recognition, they aim to gain deeper insights into the drivers of demand and assess hypothetical scenarios that could inform their decision-making process. This approach would enable them to simulate the potential outcomes of strategic shifts and improve forecasting accuracy in dynamic market conditions.

Problem statement

Owens Corning’s Insulation S&OP team seeks to leverage Causal AI technology to model the cause-and-effect relationships that drive demand for insulation materials. By understanding these causal factors, the team aims to go beyond correlation-based forecasting and build a more explanatory demand model. This model will then be used to simulate counterfactual scenarios, allowing the team to assess the potential impact of strategic decisions and external changes.

We carried out a 15 weeks of joint exercise with the Owens Corning team to deliver the Proof of Value (POV

) The POV phase is executed in four steps: Environment setup, Machine teaching, Output review, Result presentation. Environment Setup, data & knowledge collection

Machine teaching, Hypotheses Framework

Demand Forecasting Agent

Development & Integration

Result presentation

Solution dev Solution Dev

POV

Outcomes

3 Weeks

6 Weeks

1 Week

Cloud environment with TRAIN pipeline

Combined causal graph, causal analysis ready pipeline

Demand Forecasting Agent

(Backtesting for 2024 demand and forecasting for 2025 demand)

POV result

✓ Define Scope ✓ Identify resource

✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

Data preprocessing and synchronization

✓ Develop Agent (back-testing) ✓ Integrate with causal analysis ✓ Procure 2025 data for forecasting

✓ Generate POV

✓ Generate combined causal model

✓ Next step

requirement

✓ Provision cloud environment

✓ Identify SMEs, setup

Activities

5 Weeks

interviews

✓ Identify & review

knowledge artifacts in TRAIN

✓ Identify & review data artifacts

Ingest knowledge

Ingest data

model

Generate data model Built domain corpus

and intervention analysis(updated data for 2025)

Conduct interviews Generate causal hypotheses Generate combined causal model (PCM, RCM, SCM)

✓ Perform causal intervention ✓ SME review

Deliverable

collection & review

Activities dependent on the long-term scope

Depends on long-term scope (e.g., causal hypotheses, what-if simulation etc.)

report

generation discussion

✓ Configure additional agents: ✓ 2025 forecast ✓ Include Pre and post covid model

✓ Include only post covid model

✓ Data artifacts collection

Implement long-term scope (e.g., causal hypotheses, what-if simulation etc.)

✓ Report value

✓ SME review ✓ Agreed scope ✓ Knowledge artifacts

Out of Scope

✓ Data and knowledge ingestion ✓ Multi-layer causal model ✓ Causal analysis ready pipeline

✓ Working Agent ✓ Agent recommendations

✓ POV report

& review

✓ SME interviews setup

15 Weeks Kickoff

Private and Confidential 6

Completion

✓ Additional tasks added to the original plan based on OC’s suggestion to regenerate models separately for post-covid data, and to create 2025 demand forecast along with 2024 back-testing outcome

Legend:

➢ Not started ❖ In progress ✓ Complete

The Causal AI platform, TRAIN, combines expert knowledge and data to deliver forecast

Experts’ knowledge

TRAIN

TRAIN Causal AI platform

Causal Analysis

Public data

Validation I Testing hypothesis

Automated Causal Model Generation

Causal Solution Workbench

Demand forecasting agent

Application Agent

• Improved accuracy • Increased transparency • Explainable predictions

Recommendations

Proprietary data

The first enterprise causal AI platform that delivers game-changing Cost, Service, and EHS (Environmental, Health, and Safety) results. 7

The comprehensive causal model generated by TRAIN guides the agent to make an accurate demand forecast

Digital Knowledge Repository

Proprietary

Public

Government & Regulatory Reports TRAIN

TRAIN Causal AI platform

Market research & analysis reports Internal Reports & Studies

Expert interviews

Principle causal model (PCM)

Rationale causal model (RCM)

Hypothesis formulation Causal Analysis Validation I Testing hypothesis

Relevant Data Repository Economic Indicators

Structural causal model (SCM)

Public

Add con text to data

Simulations (scenarios) Recommendations

Regulations, Codes & Policy

Climate & Weather data

Proprietary

Construction Activity & Spending

Market Intelligence Data Operational & Supply Chain Data

Automated Causal Model Generation

Causal Solution Workbench

Demand forecasting agent

Application Agent

• Improved accuracy • Increased transparency • Explainable predictions

TRAIN – Project configuration, knowledge building and SME interview process

Unified project dashboard enabled problem owners, SMEs, and the POV execution team to collectively supervise the process

To be changed

The problem owner provided TRAIN with the background and problem statement as initial input for training process

TRAIN used the defined problem statement to recommend relevant knowledge categories, which SMEs then populated with documents

Using the user provided knowledge and initial problem statement TRAIN recommended a set of KPIs that the domain experts validated subsequently

To be changed

TRAIN identified relevant SME roles based on the problem statement and validated KPIs

TRAIN also generated identified areas where knowledge input is required – Parabole team led the process to capture the SME's knowledge

1. SME role generation Based on the v alidated KPIs and ing ested know ledg e artifacts, p latfo rm automatically suggests SME roles who can add value in the causal learning process

2. SME questionnaire generation Once the SME roles are valid ated, the p latfo rm produces a list of questio ns for eac h SME role

4. Conduction of interviews

3. Questionnaire summarization The POV team enabled a structured d iscussio n during interviews by summarizing the questions into spec ific c ategories, which the interviewers can use as part of the interv iew

5. Transcript processing and ingestion The transcripts from the interviews are pre-proc essed, and the upd ated transcript is used by the platform in causal learning.

TRAIN generated role-specific questionnaires for SME interviews

Post conducting the interviews, we ingested SME interview transcripts for model learning and refinement

To be changed

Model outcomes from PCM and RCM

We curated a repository of internal and external artifacts for causal learning

To be changed

TRAIN generated PCM and RCM models from ingested knowledge artifacts

As a first step, TRAIN algorithmically discovered a set of hypotheses from the ingested knowledge artifacts

Then, TRAIN built a traceable causal graph from validated hypotheses

TRAIN used its unique third variable analysis capability to handle the causal biases in the process

AI investigated the hypotheses list associated with each cause-effect relations through a systematic workflow to calculate the causal expectant score considering different causal biases: confounder, collider, mediator

Confounder bias (both cause and effects are driven by another common cause) - Backdoor analysis

Intervention analysis Another cause significantly drives the cause towards the cause Instrumental variable analysis

The causal biases are experimented using different intervention analysis Mediator bias drives the cause to effect path - Front door analysis

TRAIN identified confounders and mediators impacting KPI drivers

Coverage: TRAIN platform automatically discovered a total of ~3900 hypotheses from 38 knowledge sources

External documents • •

Total knowledge artifacts: 27 2917 hypotheses generated

Internal reports / SME Interviews •

• PCM

Total knowledge artifacts: • 9 Internal documents • 2 interviews covering 3 SME roles 948 hypotheses generated RCM

Data input and CCM generation

In the process, TRAIN defined the data needs from the Critical causal factors(from PCM and RCM) impacting the KPIs

Categories: Data availability

Original list of factors identified from PCM/RCM = 192

Final curated list of factors = 121

Sociopolitical / Global Factors Market & Product Dynamics

Data demand

Energy & Efficiency Building Characteristics

Environmental & Climate

Retrofitting & Remodeling Manufacturing & Industry

Macroeconomic Conditions

Regulations, Codes & Policy

Available

Not available

Assumed less relevant

Construction Activity & Housing 0

Note: Factors marked as ‘Assumed less relevant’ were identified based on SME input, as they do not impact the in-scope Batts & Rolls product category (e.g., industrial or non-residential factors)

Available

Not available

Assumed less relevant

Based on the data needs, we ingested the structured company data to perform causal analysis

The ingested data is used by TRAIN to automatically build the Structural Causal Model (SCM

)

TRAIN then combined knowledge and data models into a Combined Causal Model (CCM

)

Intervention analysis is then performed using CCM and the estimated causal scores to validate the hypotheses

Based on the Causal scores, we rank-ordered the set of causal factors impacting insulation demand

…

Causal factors influencing insulation demand

Total Residential Spending

Hail Public Residential Spending

Single Family Privately Owned Housing Units under construction

HMI

Total Privately Owned Housing Units under construction

Single Family Privately Owned Housing Completions

Total Privately Owned Housing Starts

RMI

Disposable Personal Income

Single Family Existing Home Sales

NonResident ial Spending on Office and Retail Multi Family Privately Owned Housing…

Singl…

Mul…

…which we categorized in four major buckets that drive insulation demand

Construction spending 25%

Construction activity 50% Economic indicators 19% Extreme weather events 6%

Construction activity

Extreme weather events

Economic indicators

Construction spending

Demand Forecasting Agent

As part of the POV exercise, we configured a ‘Demand Forecasting Agent’ to simulate demand scenarios

The agent allows users to simulate demand by configuring custom inputs

In a back-tested scenario on 2024 demand, the causal prediction is ~36% better than a statistics-based model

35.5% reduction in error (MAPE)

Applicability of POV causal model for 2025 demand forecast

While forecasting demand for 2025, we faced significant limitations given the POV nature of this work

Selection of inputs for 2025: While the 2024 back-tested model used 74 factors, for 2025 we focused only on 33 factors identified as causally relevant based on 2024 model. These were intentionally selected and sourced where available—some directly from original datasets, others estimated using 2023–2024 trends from internal or third-party sources. In a real-world implementation, these 33 drivers would also be forecasted using the causal agent.

2. Data period and comparability: The comparison dataset includes actuals for January–March 2025 and business forecasts for April–December 2025. Since the forecasted portion reflects knowledge of early 2025 performance, it isn't a fully independent forward-looking baseline. In contrast, the causal model was built solely using data available up to December 2024. 3. Differences in variable inclusion: The comparison (Actual/Business Forecast) likely includes variables such as warehouse inventory, tariffs, and other operational inputs, which TRAIN platform also needed in PCM/RCM. However, due to lack of data, these variables could not be incorporated in TRAIN. This discrepancy may impact the fairness and accuracy of direct model-to-model comparison.

Agent used the causal knowledge for 2021-2024 to forecast 2025 Q2-Q4 demand, and compared to business forecast

Scenario: Seasonal Forecast April-December 2025 using Jan 2021 to March 2025 Date Metrics MAE MAPE-vs Actual RMSE MAPE IBP vs Actual

With causal Without causal 9,541.01 29,663.92 5.42% 16.09% 12,804.49 30,755.67 9.79%

Without With causal causal

Business forecast

IBP Forecast

4/1/25

198,185.06

212,180.55

188,854.00

196,844.94

5/1/25

191,925.86

201,088.50

164,675.00

198,684.32

6/1/25

189,308.80

209,256.48

173,413.00

198,799.42

7/1/25

192,651.92

221,395.50

174,687.00

200,339.44

8/1/25

199,482.34

223,615.59

193,174.00

215,529.35

9/1/25

203,961.23

227,021.33

202,635.00

216,096.49

10/1/25

202,049.53

226,221.20

201,339.00

209,613.72

11/1/25

197,376.33

221,173.69

202,215.00

209,613.72

12/1/25

191,916.34

220,173.42

194,159.00

209,613.72

The causal forecast outperformed IBP and showed potential for improving IBP modeling using causal knowledge

Seasonal Forecast April-December 2025 using Jan 2021 to March 2025 240,000.00 230,000.00 220,000.00

210,000.00 200,000.00

190,000.00 180,000.00 170,000.00

160,000.00 150,000.00 4/1/25

5/1/25

6/1/25

With causal

7/1/25

Without causal

8/1/25

9/1/25

Business Forecast

10/1/25

IBP Forecast

11/1/25

12/1/25

The POV demonstrated using causal reduced error by ~36% while enabling transparency and defensibility in forecasting

Improved forecast accuracy using causal approach

Append causal knowledge to existing data modeling technique to improve the forecast

Ability to handle interdependencies with large number of variables

Deliver higher accuracy in forecasting even with minimal data

Provide transparency and explainability to forecasting decisions

Reduce reliance on historical data patterns, by supporting with institutional /SME knowledge

We identified value-driven opportunities to extend causal AI adoption across functions

Leverage the successful POV outcome to explore broader use of causal AI in demand forecasting, focusing on extending the model to SKU-level planning by incorporating supply, inventory, and sales inputs.

2. Build on the modular, explainable nature of the causal model to scale its use across additional product categories and operational decision points. 3. Collaborate with Owens Corning’s Manufacturing team to apply the platform in identifying causal drivers of material

quality issues, enabling proactive and explainable root-cause analysis.

1100 Cornwall Road Suite 214 Princeton Corporate Center Monmouth Junction, NJ 08852 parabole.ai

About Parabole.ai At Parabole, we offer the first enterprise-ready Causal AI Platform. Our software enables organizations to create and use causal models to solve pressing decision problems within Manufacturing, CPG and Oil & Gas industry . Our customers use our software to discover, methodically, the “Why” of a problem and solve it with context-based variable optimizations.

44 44

Our roadmap

Faster, natural

Big compute

Causal analysis

Slow, legacy

Optimal compute Statistical analysis Amount of data

Compute landscape (2025-2027

)

Compute landscape (2028 - onward

)

About Parabole.ai

Who we are

enterprise causal AI platform for manufacturing, chemicals, consumer products and oil & gas

proprietary AI agents targeting supply chain, process control, asset management and other critical business functions

10+

Fortune 500 are clients

Recognized as a leader in causal AI by

Gartner

175+ years

combined digital transformation experience of the founders and board advisors

30+

engineers & data scientists from leading schools

Parabole ai expertise • Parabole.ai’s TRAIN platform creates tangible financial impact by standardizing, improving and accelerating complex operational decision-making with causal AI • TRAIN is highly interoperable with standard enterprise systems and does not require complex IT/OT integrations, allowing for rapid implementation and scaling

AI Comparison: Gen AI, Predictive AI and Causal AI

Gen AI: Applies ML, LLM’s and RAG’s to generate images, text, videos, and other media in response to prompts. Predictive AI: Aggregates and maps all data, then uses ML and predictive statistics to create correlation relationships between variables and outcomes. Predicts outcomes based on correlation. Causal AI: Aggregates relevant data and knowledge then applies Causal Inference statistics to define the causal relationships between variables and outcomes. Predicts outcomes based on causation.

Gen AI

Causal AI

Predictive AI

Scale

Predictive AI

Decision impact

All three should play a role in your value creation plan

Causal Inference

Causal AI and Parabole’s Platform provide significant advantages over predictive solutions

Moving from Correlation to Causation delivers superior

• Efficacy: Causation, not correlation

Exponential acceleration of

• Explainability: Documented logic trail

CI/Optimization Efforts

• Adaptability: First principles-based

• Integrates with and automates existing CI

Supply chain planning: Raw materials procurement plan optimization

• Efficiency: Only relevant data and knowledge • Response Time

Methodologies (LSS) Gartner 2024 Hype Cycle Report

• Continuous learning platform • CI projects ingested as inputs

Parabole’s Causal AI Platform • Operated by SME’s and data scientists

• “Self Service” Platform enables

• Can be deployed on any process

• RAPID replication capability

• Faster and at a lower cost, the deployment

• SME Multiplier

• Causal Knowledge Repository provides big added value

• Full IP Protection

Causal AI is the Next Generation of problem-solving capability

1100 Cornwall Road Suite 214 Princeton Corporate Center Monmouth Junction, NJ 08852 parabole.ai

50 50

Topics: causal-ai, software-as-intentOpen in the Radiant ↗All dispatches