Predictive Analysis

Every predictive engagement begins with a written decision specification. What is the decision the model will inform? Who makes it? What is the cost of a miss in either direction? How quickly does the decision-maker need the prediction, and at what level of granularity? What is the current process being augmented or replaced, and how good is it on the same metric? Without these answers, modelling work tends toward whatever metric the data scientist is most comfortable optimising; with them, every later trade-off has a reference point.

We are particularly attentive to asymmetric costs. In demand forecasting, an over-forecast on perishable stock costs differently from an under-forecast on a flagship product; in churn modelling, a false positive may trigger an expensive retention offer while a false negative loses a customer altogether. We design loss functions, decision thresholds and confidence requirements to reflect those asymmetries explicitly, rather than relying on a default symmetric metric to flatter the model.

Horizon is treated with similar care. A short-horizon forecast and a long-horizon forecast are usually different problems, requiring different features and different models. We resist the temptation to pretend that one model covers all horizons, and we build the planning interface to expose horizon explicitly.

• Decision specification before modelling
• Asymmetric loss and threshold tuning
• Horizon-appropriate model selection
• Hierarchical structures (region, channel, SKU)
• Cold-start strategies for new entities
• Override workflows for human judgement

Our practice draws across the full predictive toolkit: classical exponential smoothing and ARIMA where simplicity wins; gradient-boosted trees with rich feature engineering for tabular problems; state-space models and Bayesian structural time-series for interpretable, probabilistic forecasts; modern global deep models such as Temporal Fusion Transformers, N-BEATS and DeepAR where cross-series learning helps; survival models for time-to-event problems; and isolation forests, autoencoders and probabilistic methods for anomaly detection. Each has a domain in which it is the right answer.

We are particularly fond of probabilistic forecasting. A point forecast forces decision-makers to act as if the future were certain; a probabilistic forecast lets them price risk explicitly. Quantile forecasts, prediction intervals and full predictive distributions enable inventory policies, capacity decisions and pricing strategies that point forecasts cannot support. The additional engineering effort is modest, and the operational benefit is large.

We also pay close attention to hierarchical reconciliation. Forecasts produced independently at different levels of a business hierarchy — country, region, store; brand, category, SKU — are almost never coherent with one another. Reconciliation methods (MinT, top-down, bottom-up combinations) restore coherence and almost always improve accuracy at the same time.

“A well-calibrated probabilistic forecast is worth more than a slightly-more-accurate point forecast.”

Most predictive projects fail their first contact with production not because the model is bad, but because evaluation was carried out on a metric that did not represent the operational cost of being wrong. We design evaluation harnesses that mirror the deployment setting: rolling-origin cross-validation that respects time, hold-out periods that include the seasonal patterns the model will face, and metrics that incorporate the asymmetric costs identified during framing.

We backtest aggressively, including against the deliberate baseline of doing whatever the business currently does. A new model that beats a sophisticated baseline by a small margin is far more impressive — and far more reliable — than one that beats a strawman by a large one. Where possible, we recreate historical decisions and measure the dollar cost of the model's prediction against the cost of the decision actually taken at the time.

Calibration receives explicit attention. A probabilistic forecast that promises 90% intervals must contain the truth roughly 90% of the time, or the planning that depends on it will quietly mis-fire. We test calibration on hold-out data, and we recalibrate where necessary using techniques such as conformal prediction.

A forecast that lives in a separate dashboard is rarely acted on. We integrate predictive systems directly into the planning workflows they are meant to inform — into the S&OP cycle, the demand-planning tool, the CRM, the contact-centre console — so that the forecast is consumed where the decision is made. This integration work is typically the difference between a model that is admired and one that is used.

Human override is treated as a feature, not a sign of failure. Planners hold context the model does not — a competitor's launch, a known supply disruption, a marketing decision not yet visible in the data — and a system that cannot accept their input will quickly be ignored. We design override interfaces that capture the reason for the override alongside the value, so that overrides themselves become training data for the next cycle.

Operational monitoring covers prediction error in production, drift in input distributions, and changes in the bias and calibration of the forecasts. Where models retrain on a schedule, the schedule is informed by the rate of change observed; where they retrain on signal, the trigger conditions are explicit and reviewed regularly.

Improved predictive accuracy translates into business value through specific operational mechanisms — lower buffer stock, higher service levels, fewer stockouts, better-staffed contact centres, more selective retention spend, more confident capacity decisions. We work with finance and operations to make those mechanisms explicit before the project starts and to measure them after deployment, so that the value of the system can be defended in any future business case.

Across our engagements, the most common single outcome is a meaningful reduction in working capital tied up in inventory or in conservative buffers, alongside an improvement in service levels. The second most common is the retirement of a manual forecasting process whose effort was previously hidden in operational headcount.