Why forecast without a language model?

The forecast has to be reproducible and explainable. No language model touches the numbers; a fixed seed and a data fingerprint make every run repeatable.

What does the Monte Carlo simulation do here?

It replays attrition, spend share by tenure cohort and returns, and gives a range of outcomes drawn from real historical spread instead of fixed percentages.

How are incomplete months handled?

They are flagged and imputed with explicit, labelled assumptions. When per-performer data is missing, a team-level bootstrap is used.

How is the forecast checked against reality?

By an as-of-date back-test: the model runs on history truncated to a past date and is compared with what happened.

Is zero departures in a small sample treated as zero risk?

No. Attrition is smoothed, so a small sample with no exits does not mean no risk.

Does the tool make the decision?

No. It is an advisory tool and the decision stays with a person.

Team forecasting system: compare teams, predict the effect of roster changes

A deterministic forecasting engine: it models each performer separately, sums to the team and shows a range of outcomes from a Monte Carlo simulation. No language model touches the numbers.

§01 The case in detail

Task

Compare teams and predict how roster changes (new hires, departures) change a team’s results.

How it works

The system is a deterministic forecasting engine: no language model touches the numbers. It models each performer separately, then sums to the team.

A Monte Carlo simulation replays attrition, spend share by tenure cohort and returns, giving a range of outcomes drawn from real historical spread, not fixed percentages.

Incomplete months are flagged and imputed with explicit, labelled assumptions. Every run is versioned with a data fingerprint, and can be back-tested on history truncated to a past date.

Stack

FastAPI · server-rendered pages · e-mail one-time-code sign-in · three containers: web, build, compute (build and compute have no network)

Why no LLM

The forecast has to be reproducible and explainable. A fixed seed and a data fingerprint make every run repeatable, and the decision stays with a person: the tool is advisory.

Principles

Data
Per-performer data, not team aggregates
Tenure cohorts as a frame
Team-level bootstrap when per-performer data is missing
Unclosed months flagged
Model
Monte Carlo with a fixed seed (reproducible)
Attrition with smoothing (zero exits in a small sample is not zero risk)
Returns drawn from neighbours of similar spend size
Team budget as a shared ceiling
Control
As-of-date back-test against the fact
Risk rules (concentration, drawdown, unconfirmed spend)
Versioned runs with a data fingerprint
Advisory tool: the decision stays with a person

Architecture

Architecture as text: who passes what to whom

Sign-in

Team member ↔ CDN + proxy: requests
CDN + proxy ↔ Web: filtered
Web → Mail service: one-time code
Web ↔ Results DB: codes, sessions

Upload + build

Web → Uploads: stage Excel
Web → Results DB: queue request
Host timer → Build job: start
Uploads → Build job: read files
Results DB → Build job: requests
Build job → Team data: write
Operator → Team data: manual push

Forecast

Team data → Compute job: read only
Compute job → Results DB: runs, risks

View

Results DB → Web: runs, risks
Web → Team member: pages

§02 Deterministic Monte Carlo forecasting without an LLM: problem, approach, outcome

Problem: Compare teams and predict how roster changes (new hires, departures) change a team’s results.
Approach: A deterministic forecasting engine that models each performer separately and sums to the team. A Monte Carlo simulation replays attrition, spend share by tenure cohort and returns. Incomplete months are flagged and imputed with explicit, labelled assumptions.
Outcome: The forecast is a range of outcomes drawn from real historical spread, not fixed percentages. Every run is versioned with a data fingerprint, repeatable with a fixed seed, and can be back-tested on history truncated to a past date.

Short answers

Why forecast without a language model?: The forecast has to be reproducible and explainable. No language model touches the numbers; a fixed seed and a data fingerprint make every run repeatable.
What does the Monte Carlo simulation do here?: It replays attrition, spend share by tenure cohort and returns, and gives a range of outcomes drawn from real historical spread instead of fixed percentages.
How are incomplete months handled?: They are flagged and imputed with explicit, labelled assumptions. When per-performer data is missing, a team-level bootstrap is used.
How is the forecast checked against reality?: By an as-of-date back-test: the model runs on history truncated to a past date and is compared with what happened.
Is zero departures in a small sample treated as zero risk?: No. Attrition is smoothed, so a small sample with no exits does not mean no risk.
Does the tool make the decision?: No. It is an advisory tool and the decision stays with a person.

§03 Related

CRM for a network of martial-arts clubsWhatsApp funnels, reports, trainer pay
Approval system for financial requestsPersonal project: rules as data, encryption, audit log
Person dossier aggregator from official sourcesMany sources, one reviewed dossier
Forecasts and dashboardsNumbers computed by code, with an honest range
LLM integration into working systemsWire a model into the POS, booking, CRM, messengers

Have a task like this? Write to me.

Message on Telegram oleksandr@semishan.pro

Send me a message

It goes straight to my Telegram. I reply to the address you leave.