How Causal Discovery and Risk Simulation Enhance Demand Forecasting Reliability

When AI forecasts "demand will increase 30% tomorrow," what's the first question practitioners ask? It's "why?" They want to understand which factors drove the demand increase, whether they can safely increase orders based on this forecast, and if any unexpected variables might be at play. Causal discovery and risk simulation, core explainable AI techniques, directly address these questions. Rather than simply presenting forecast values, they provide the reasoning behind predictions and quantify uncertainty, enabling practitioners to make decisions with confidence.

‍

Finding the Real Causes Behind Demand Fluctuations with Causal Discovery

Correlation and causation are fundamentally different. Ice cream sales and shark attacks at beaches both increase during summer, but there's no direct causal relationship between them. Rising temperatures, a common cause, trigger both phenomena independently. However, typical data analysis can mislead us into mistaking spurious correlations for true causes. If you seriously considered the correlation between shark attacks and ice cream sales, you might implement a misguided shark warning system.

Causal discovery models identify true causes from data. They can uncover likely causal structures such as rising temperatures increasing beach visits, which in turn drive ice cream sales. Simultaneously, they distinguish that shark attacks occur independently due to temperature changes and have no direct causal relationship with ice cream sales.

Finding Root Causes in Call Center Problems

Consider a real business situation. A company experiences increased customer churn. Simultaneously, call wait times lengthen, problem reports increase, discount coupon distribution rises, and call abandonment rates climb. On the surface, multiple problems seem to occur concurrently.

Causal discovery models can reveal the actual structure of this complex situation. For instance, demand increases might have lengthened wait times, which led to increased call abandonment and ultimately customer churn. Longer wait times also triggered more problem reports, prompting increased discount coupon distribution. While discount coupons partially reduced churn, they weren't a fundamental solution.

Understanding this causal structure enables effective countermeasures. Since the root cause is insufficient staffing for increased demand, hiring more representatives to reduce wait times becomes the direct solution. Discount coupons should only serve as a temporary measure to alleviate symptoms. This prevents the flawed conclusion that "just increase discounts."

Developing Strategy Through Forecast Cause Analysis

When AI forecasts "next month's sales will increase 30%," how much should practitioners trust this prediction? Causal analysis decomposes the factors contributing to the forecast. You discover that increased TV advertising contributed 40% of the impact, seasonal effects accounted for 30%, a competitor's quality issues contributed 20%, and unpredictable social media viral effects made up 10%.

This analysis empowers practitioners to make informed decisions. TV advertising deserves continued investment. Seasonal effects are automatically recurring patterns requiring no special response. However, relying on competitor problems is risky since this effect disappears once competitors resolve their issues. Social media virality should be considered only as an unpredictable bonus.

How Causal Discovery Works

Causal discovery operates through three stages. First, it maps the relationship network among variables in the data. It visualizes structures like temperature affecting beach visits and shark attacks, while beach visits influence ice cream sales.

Second, statistical tests separate true causal relationships from coincidental correlations. When controlling for beach visits, does temperature directly affect ice cream sales? If the answer is no, only indirect effects exist. When controlling for temperature, do shark attacks affect ice cream sales? Again, if the answer is no, it confirms a spurious correlation.

Third, it determines the direction of causation. It clarifies which side is the cause and which is the effect. This tells practitioners where to intervene most effectively.

Precisely Diagnosing Causes of Sudden Changes

Imagine a situation where demand suddenly plummets. Website traffic drops 30%, search ad clicks decrease 25%, mobile app usage declines 10%, but customer inquiries actually surge 50%. Multiple metrics deteriorate simultaneously, but what's the root cause?

Causal discovery reveals the actual causal structure. Degraded website loading speed was the root cause. This triggered traffic reduction, which led to decreased search ad efficiency. Some customers attempted to migrate to mobile, causing a surge in inquiries during this process. The solution is clear: resolve the website technical issues to eliminate the root cause.

Causal discovery models aren't perfect in every situation. You might not be able to measure all factors, and domain expert knowledge is essential. Never blindly trust results without verification. Sufficient data volume is required, and measurement errors can distort outcomes. However, when you recognize these limitations and use them appropriately, causal discovery becomes a powerful tool for correct decision-making.

‍

Managing Uncertainty Through Risk Simulation for Decision Support

Risk simulation is a virtual laboratory that answers "what if?" questions. You can test thousands of scenarios before actually executing them. Before launching a new product, changing your supply chain, or modifying pricing policies, you can experience various possibilities in advance.

Assessing New Product Launch Risks

Suppose you're launching a new smartphone model. Demand forecasts range from 500,000 to 2 million units, component prices may fluctuate plus or minus 20% from current levels, exchange rates could vary plus or minus 10%, and there's a 50% probability competitors will cut prices.

Traditional analysis presents only an average scenario. It delivers a single prediction of 1.25 million units in demand with 10 billion won in profit. However, the average scenario almost never actually occurs. Risk simulation runs 10,000 virtual scenarios. Results appear as distributions. There's a 30% probability of profits between 5 and 10 billion won, 40% probability between 10 and 15 billion won, 20% probability between 15 and 20 billion won, and a 10% probability of losses.

The key finding is that worst-case scenarios could result in 5 billion won in losses. The most likely outcome is 12 billion won in profit, with only a 5% chance of exceeding 20 billion won in revenue. With this information, practitioners can secure emergency funds for potential 5 billion won losses and adopt a strategy of conservatively producing only 1 million units initially while keeping additional production capacity on standby.

Preparing for Supply Chain Disruption Scenarios

Consider a situation with only one critical component supplier. What happens when you simulate a natural disaster causing a one-month supply disruption? Week one shows no impact as inventory covers demand. Week two sees production decline 30%, week three drops 70%, and by week four production completely halts. Total revenue loss reaches 30 billion won with severe brand image damage.

Based on these simulation results, you can develop countermeasures. Increasing inventory from two to four weeks costs 1 billion won annually, while contracting with an alternative supplier costs 500 million won annually. Establishing an emergency production conversion plan is also necessary. Ultimately, investing 1.5 billion won annually hedges against 30 billion won in risk.

Predicting Subscription Price Change Impacts

When considering a 10% subscription price increase, customer churn is expected to range from 5% to 20%, new customer acquisition could vary from minus 30% to plus 10%, competitors might run promotions, and customer lifetime value will change.

Running 10,000 Monte Carlo simulations reveals result distributions. The worst case has a 5% probability with 20% churn, minus 30% new acquisition, minus 15% revenue after one year, and minus 300% ROI. The expected scenario has a 60% probability with 8% churn, minus 10% new acquisition, plus 5% revenue after one year, and plus 50% ROI. The best case has a 15% probability.

In conclusion, positive outcomes are expected with 85% probability. Based on this information, you can execute the increase but take a phased approach while simultaneously running churn prevention programs and new customer acquisition campaigns.

How Risk Simulation Works

Risk simulation operates through four stages. First, create a digital replica of the actual business. For a manufacturing process, include five production lines in your digital model with each line's hourly throughput, quality defect rates, worker shift times, machine failure probability, and inventory storage capacity.

Second, set scenarios you want to test. Define normal operations, doubled demand, one machine breakdown, 20% worker absence, raw material price spikes, and combined crisis situations where multiple problems occur simultaneously.

Third, simulate each scenario thousands of times while varying random variables. In the first run, demand increases 1.8x with 2.1% defect rate and no breakdowns. In the second run, demand increases 2.2x with 1.9% defect rate and one breakdown occurs. Repeat this 10,000 times.

Fourth, analyze results. You discover 90% probability of meeting delivery deadlines, 5% probability of one-week delays, 3% probability of two-week delays, and 2% probability of serious disruptions. You identify line 3 as the bottleneck and find the most dangerous scenario combines demand surges with machine breakdowns.

Experimenting Without Cost and Understanding Complex Situations

Risk simulation's greatest value is enabling cost-free experimentation. When a new logistics center requires 50 billion won investment and cannot be reversed once built, you can simulate three candidate sites. The urban location costs 60 billion won construction with 2-hour average delivery time and 120 billion won total 10-year cost. The suburban location costs 40 billion won construction with 4-hour delivery time and 100 billion won total 10-year cost. The middle location costs 50 billion won construction with 3-hour delivery time and 95 billion won total 10-year cost. Through simulation, you select the optimal balance point of the middle location and discover a 25 billion won savings opportunity before actual investment.

You can also understand complex situations like hospital emergency room congestion. Patient arrivals are random and unpredictable, treatment times vary by patient, medical staff work in shifts, bed availability depends on admissions and discharges, and priorities differ by emergency level. Testing several alternatives through simulation reveals that adding two doctors reduces wait times 15% but costs 300 million won annually. Establishing a fast-track treatment room reduces wait times 35% with 200 million won annual cost. Implementing a reservation system reduces wait times 10% with 100 million won annual cost. You can choose the best value option of establishing the fast-track treatment room.

‍

ImpactivAI Deepflow's Explainable AI Technology

ImpactivAI's Deepflow is a solution that applies causal discovery and risk analysis to actual demand forecasting. Utilizing 224 advanced machine learning and deep learning models, it not only generates highly accurate forecast values but also clearly explains the reasoning behind forecast results.

Deepflow quantitatively presents external variables that influenced model predictions, such as macroeconomic indicators and industry attribute data, along with each variable's contribution rate. By providing brief explanations for each external variable and the top 20 contribution rates, practitioners can immediately grasp which factors most significantly impact current forecasts.

The recently added LLM-based analysis feature in Deepflow has advanced explainability to the next level. Leveraging generative AI, it automatically generates past sales trend and seasonal pattern analysis, future demand outlook with forecast reasoning, and customized execution strategies by department. Reports systematically present key risk and opportunity factors for each practitioner role in sales, marketing, and SCM, and provide action plans optimized for each department. Practitioners save time previously spent on complex data interpretation and report writing, allowing them to focus solely on strategic decision-making.

The BI dashboard provides at-a-glance visibility of stockout or excess SKUs and automatically calculates days of inventory on hand and optimal production quantities reflecting future sales changes. Companies reduce unnecessary costs and stockout risks in inventory management, while practitioners can immediately grasp the situation through visualized materials without additional editing. The MI dashboard provides external environment data including exchange rates, interest rates, raw material prices, and global economic indicators, and supports proactive response to market volatility through 3-month short-term AI forecasts for exchange rates and oil prices.

How Should You Use Causal Discovery and Risk Simulation?

Integrating Both Techniques

Using causal discovery and risk simulation together becomes even more powerful. Consider marketing budget optimization as an example. First, understand the structure through causal analysis. TV advertising increases brand awareness, awareness leads to website visits, and visits convert to purchases. Social media advertising bypasses the awareness stage and connects directly to purchases. Search advertising doesn't create new demand but merely converts existing demand.

Based on this causal structure, find optimal allocation through simulation. After 10,000 simulations, the current allocation of TV 40%, social media 30%, search 30% yields average 150% ROI but with high volatility. The optimal allocation of TV 50%, social media 40%, search 10% yields average 180% ROI with lower volatility. Reducing search advertising while increasing TV and social media is recommended.

Finally, assess risks. TV advertising effects can vary plus or minus 20%, social media algorithm change risks exist, and seasonal factors must be considered. When all scenarios are considered, the optimal allocation still shows superior performance, making it highly executable.

Practical Implementation Guide

When adopting causal discovery, first define clear questions. Specifically ask why sales declined or which factors influence churn. You need at least hundreds to thousands of observation data points, measuring change data over time and related variables. Collaborating with domain experts to verify and interpret results is crucial. Start with simple problems to accumulate success cases and spread adoption throughout the organization.

When implementing risk simulation, clarify simulation objectives. Define which decisions to support and which risks to evaluate. Identify key variables that are most uncertain and most impactful. Set realistic scenarios including extreme situations based on historical data. Ensure simulation results lead to decisions and response plan development by deriving actionable insights.

Causal discovery and risk simulation are core technologies that elevate AI demand forecasting beyond simply presenting numbers into a practical decision-making tool. Financially, they avoid misguided investments to reduce costs, select effective strategies to increase profits, and prevent unexpected losses to mitigate risks.

Temporally, they enable fast decision-making with data-driven confidence, reduce trial and error through virtual testing, and rapidly respond to changes to capture opportunities. Strategically, they secure competitive advantage through more accurate forecasts, enable innovation in safe experimental environments, and build trust through transparent decision-making rationale.

Only when you understand why forecasts turned out as they did, manage risks in uncertain situations, and secure concrete execution strategies can AI forecasts truly convert into business value. Explainable AI is no longer optional but essential.