I
Customer records and pipeline
A CRM your sales team opens every morning.
See the CRM build
Causal64: machine learning and forecasting for companies and public bodies
The prediction methods of quantitative finance have stayed inside trading desks. We bring them to companies and public bodies, to improve the results that depend on a number. Every model we ship is tested on data it has never seen.
What we do
Forecasting and statisticsDemand, load and price forecasts, studies, surveysExplore
Machine learning and AIScoring, classification, text, risk modelsExplore
Data and CRMPipelines, clean customer data, reportingExplore
SoftwareCustom applications and internal toolsExplore
PlatformsDashboards, client portals, multi-user systemsExplore
Software we build
II
Model and simulation studio
Data, pipeline, runs and the model’s behaviour in one workspace.
See the model studio buildIII
Demand planning software
History, statistical forecast and consensus in one grid, with the alerts.
See the demand planning buildWherever a decision rests on a number.
Benchmarks
How many claims each motor policy will file, on the public benchmark actuaries use.
Poisson deviance ×100
23.5024.0024.5025.0025.50
Lower is better. The axis starts at 23.2. The other bars are the scores published by their authors on the same test split. How the claims benchmark was run
Deaths from armed conflict in every country, each month, up to fourteen months ahead.
Causal64VIEWS baselinesChallenge entries
10.020.030.040.050.060.0
Every entry rescored with the official code. The first two are too close to rank. How the conflict forecast was scored
Monthly demand for 2,404 spare parts that sell rarely, against the methods planning software ships with.
| Causal64 | Moving average | TSB | Exponential smoothing | SBA | Croston | Last month | |
|---|---|---|---|---|---|---|---|
| Over- or under-forecast of total demandclosest to 0 is better | +6% | +18% | +21% | +23% | +16% | +37% | -10% |
| Error on irregular, lumpy partsagainst exponential smoothing, lower is better | -3.6% | -1.3% | -1.1% | ref. | -0.7% | +3.8% | +16.7% |
| Error on all 2,404 partsRMSSE, lower is better | 0.574 | 0.557 | 0.561 | 0.562 | 0.590 | 0.615 | 0.618 |
Every setting frozen before the six test months were scored. Best figure of each row in bold; on lumpy parts the gap with the moving average is within noise. The full demand forecasting test
Research
Research noteHow we treat dataWhere each number comes from, what was known on the day, and how we catch data that leaks the answer.Read the note
Research noteOur gatesThe tests a result must pass before we report it, and why a sealed sample is read only once.Read the note
Research noteDetecting overfittingHow we tell a real effect from the best of many lucky tries, with the cases that fooled us first.Read the note
More notes: Credit model validation, a worked report Time series cross-validation, measured Our benchmark record Intermittent demand forecasting, measured Small area estimation, checked against a census
Team
We ship with
- Supabase
- Cloudflare
- Hetzner
- Anthropic
- Nebius
Names are trademarks of their owners, used to describe the services we build with. No partnership is implied. Trademarks
Send us the brief.
A tender, a product to build, a dataset or a decision that needs a number. Tell us what you need. We reply with how we would do it and what it would cost.
contact@causal64.com