Forecasts you can check against what actually happened.

Where the problem shows up

Most businesses begin predictive work with the model. A team hears that forecasting can improve planning, chooses a tool or an algorithm, and feeds it whatever data is available. A model produces an answer quickly, and an answer of any kind feels like progress.

A prediction can be no better than the history it learns from. If the history is inconsistent between sources, if definitions changed over the years, or if two analysts prepared the dataset differently, the model learns those inconsistencies along with the pattern. The output looks precise while resting on something nobody has verified, and budgets, staffing and inventory are planned against a number of unknown reliability. What happens to confidence in forecasting after the first large miss, when nobody can say whether the model, the data or the market was at fault? And how can a business improve a prediction when it cannot reproduce the dataset that produced it?

The better principle is to start from the decision, settle and document the data underneath it, begin with a simple baseline, and compare every forecast with what happened. That is the work I do.

What I have changed for others

Market research data: a reproducible base for testing ideas

A quantitative research setup prepared its own copy of the data for each experiment, so one result could not be compared reliably with another. I designed a PostgreSQL environment that holds more than 6.4 million historical price records, with 21 native timeframes kept as recorded and raw data separated from derived research outputs. It also has dedicated structures for research runs and detected market events, so experiments can be repeated and tested against the same data. Read the case study.

100+ people: forward-looking targets in a weekly report

Planning in a large organisation lived in scattered messages, so individual targets and progress were hard to see. I built a system that produces a Weekly Trajectory Report for each employee, with planned targets, progress achieved and next targets, and a company-level summary. More than 100 people now work from one recurring forward-looking view. Read the case study.

I do not publish forecast accuracy figures that have not been measured on a client's own data.

What changes for your business

You know whether your history is good enough to forecast from before any money is spent on a model. Forecasts run on a schedule, and each one is compared with the actual outcome, so you learn how much weight to give it. The work can include a data audit, a reproducible dataset with documented preparation, baseline models, backtesting on held-out data, scheduled forecasting workflows and reports that compare forecast with outcome.

Where this leads next

Prediction rests on the layers beneath it, so most engagements begin with information systems or data liquidity before a model is built. If your data is already consistent and reproducible, we can begin with a baseline.