The Problem with Pure AI Forecasting
Feed your sales data into a machine learning model and it will find patterns. That's what ML does. The trouble is, it will find patterns whether they exist or not.
A neural network given enough parameters will happily fit your historical data with near-perfect accuracy—and then produce forecasts that bear no relationship to reality. It learned the noise, not the signal. When that "highly accurate" model predicts December sales of £100K and you actually sell £200K, the consequences are real: stockouts, missed revenue, damaged customer relationships.
This is the fundamental challenge with AI-driven demand forecasting. The models are powerful, but they're also dangerous. They'll confidently tell you exactly the wrong answer.
Our Approach: AI Variables, Statistical Validation
Ordeen Forecasting takes a different path. We use AI to augment your data—discovering external signals you'd never think to include—but we validate every variable through rigorous statistical testing before it touches your forecast.
The result: models that are genuinely predictive, not just historically accurate. Models you can explain to your finance director. Models that tell you why December sales spike, not just that they do.
How It Works
Step 1: Data Enrichment
Your sales history is just the starting point. We augment it with:
Lagged variables — Last month's sales, last quarter's trend
Transformed variables — Growth rates, moving averages, seasonal indices
External data — Google Trends for your product category, macroeconomic indicators from FRED, weather patterns
AI-discovered signals — Our system identifies external factors correlated with your demand that humans might miss
A typical forecasting project might start with 50+ candidate variables. Most of them are noise. The statistical machinery separates signal from noise.
Step 2: Four-Stage Variable Selection
This is where econometrics earns its keep. Every candidate variable passes through four statistical gates:
Stage 1: Correlation Screening We test each variable against your target using Pearson correlation (linear relationships), Spearman correlation (monotonic relationships), and mutual information (non-linear relationships). If a variable shows no meaningful relationship by any measure, it's eliminated.
Stage 2: Multicollinearity Removal Variables that measure the same underlying phenomenon create unstable models. Using Variance Inflation Factor (VIF) analysis, we identify and remove redundant variables while keeping the most predictive one from each correlated group.
Stage 3: Endogeneity Detection Does marketing drive sales, or do high sales periods trigger marketing spend? Reverse causality produces misleading coefficients. We apply Durbin-Wu-Hausman and Granger causality tests to flag variables where the causal direction is ambiguous.
Stage 4: Interaction Effects Sometimes the magic is in combinations. Price sensitivity might only matter during promotions. We test variable pairs for significant interactions—but limit them strictly to prevent overfitting.
Step 3: Overfitting Protection
Even after rigorous variable selection, models can still overfit. We apply three automated checks:
R² ceiling — If the model explains more than 95% of historical variance, it's almost certainly fitting noise
Observation ratio — We require at least 10 data points per variable in the model
Cross-validation gap — If training accuracy exceeds cross-validated accuracy by more than 20%, the model doesn't generalise
When overfitting is detected, we automatically fall back to simpler time-series methods. A slightly less "accurate" model that actually predicts beats a perfect historical fit that fails forward.
The AI-Statistics Partnership
The relationship between AI and statistics in our system is complementary, not competitive.
AI excels at:
Finding unexpected external signals
Detecting non-linear patterns
Processing unstructured data (trends, sentiment)
Generating hypotheses at scale
Statistics excels at:
Validating whether patterns are real or coincidental
Quantifying uncertainty
Preventing overfitting
Producing interpretable, defensible models
We let each do what it does best. AI proposes; statistics disposes.
A Concrete Example
Consider a building materials distributor. AI might discover that Google Trends for "home renovation" correlates with their sales—something no analyst would think to check. But correlation isn't causation. Our statistical tests examine:
Is the correlation stable across different time periods?
Does the search trend lead sales, or just move alongside them?
Does adding this variable improve out-of-sample predictions?
Does it remain significant after controlling for seasonality and economic conditions?
If the variable passes all tests, it enters the model with a quantified coefficient. If December sees a spike in "home renovation" searches, we can estimate the demand uplift—with confidence intervals.
Beyond Point Forecasts
A forecast without uncertainty bounds is a guess with delusions of precision.
Every Ordeen forecast includes:
Confidence intervals — The range within which actual demand will likely fall
Model diagnostics — R², adjusted R², cross-validation scores
Variable importance — Which factors are driving the forecast
Scenario analysis — What happens if key inputs change
This isn't just statistical box-ticking. When your S&OP process debates whether to build inventory for a promotional period, knowing that the forecast has a ±15% confidence interval versus ±40% changes the conversation entirely.
Handling Difficult Demand Patterns
Not all products behave the same way. Spare parts sell sporadically. New products have no history. Seasonal items have complex patterns.
Intermittent Demand When demand is sparse or lumpy—common in spare parts, specialty items, or long-tail SKUs—standard forecasting methods fail. We automatically detect intermittent patterns and switch to specialised methods (Croston's, SBA, TSB) that separately forecast demand occurrence and demand size.
Hierarchical Reconciliation Forecasts need to add up. The sum of SKU forecasts should equal the product group forecast, which should equal the total forecast. We apply optimal reconciliation methods (MinT) to ensure consistency across all levels while preserving the information from each level of detail.
Seasonality Detection Rather than assuming seasonality exists, we test for it. Our system identifies which months show statistically significant deviations from baseline—purely from your data, with no assumptions about Q4 spikes or summer slowdowns.
What This Means for Your Business
Accuracy you can trust Not just high R² on historical data, but genuine predictive power validated through cross-validation.
Forecasts you can explain Every prediction traces back to quantified drivers. When the board asks why you're building inventory, you have the answer.
Early warning on model degradation Continuous monitoring detects when market conditions shift and your model needs updating—before forecast errors pile up.
Integration with your workflow API-first design connects to your ERP, planning tools, and BI platforms. SAP Business One integration available.
The Bottom Line
AI is a powerful tool for demand forecasting. But power without discipline produces confident nonsense.
By combining AI's pattern-finding capability with statistical rigour, Ordeen Forecasting delivers models that are both sophisticated and trustworthy. Models that work in the real world, not just on historical data.
