forecast
Skill4 starsChanged 6 days agoMITView on GitHub
Build a time-series forecast for a business metric — baselines first, rolling-origin backtest, calibrated prediction intervals, bias check, and a monitoring plan — and report it separately from the target and the plan. Use for revenue, demand, signups, capacity, or any "what will this be next quarter" question.
8 other projects use a version of this.
Adopted by
Most recently taken up by
- mkaavish/forecastClaudeSkilltheir own version, first seen Nov 8, 2026
- anthropics/knowledge-work-pluginstheir own version, first seen Oct 13, 2026
- akachandan1/proofstacktheir own version, first seen Oct 11, 2026
- skyf0xx/gambittheir own version, first seen Sep 30, 2026
- ai-analyst-lab/ai-analysttheir own version, first seen Sep 10, 2026
- mbwsims/claude-universetheir own version, first seen Apr 14, 2026
Other versions8 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 8
- anthropics/knowledge-work-plugins★ 36ksame name
- ai-analyst-lab/ai-analyst★ 304same name
- mbwsims/claude-universe★ 71same name
- microsoft/dataverse-business-skills★ 51same name
- skyf0xx/gambit★ 29same name
- mkaavish/forecastClaudeSkill★ 1same name
- akachandan1/proofstack★ 0same name
- KevinReactiv8/claude-skills★ 0same name
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository21 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- analyzeSkill
- build-dashboardSkill
- build-modelSkill
- build-pipelineSkill
- causal-analysisSkill
- connect-dataSkill
- define-kpisSkill
- evalSkill
- experimentSkill
- health-checkSkill
- investigate-metricSkill
- make-deliverableSkill
- powerbiSkill
- privacy-reviewSkill
- research-domainSkill
- scorecardSkill
- setup-teamSkill
- tableauSkill
- triageSkill
- upgradeSkill
The file
We split this skill into the rules it gives an agent
4 rules. 0 are common, 3 are distinctive, and 1 is very specific to this project.
0
Common
Most of the 8 edited versions we read keep these.
3
Distinctive
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 10
Read analytics.md Part 6 first.
- Distinctive: line 30
Reconcile hierarchies: If the forecast is by segment or region, the parts must sum to the total.
- Distinctive: line 49
Never a point forecast without a range.
- Distinctive: line 52
When the forecast changes materially from the last one, say what moved it — new data, a changed assumption, or a model change — and by how …
“Common” is measured: kept by at least 6 in 10 of the 8 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Forecast
Owner: data-scientist, with analytics-engineer for the history and performance-monitor for tracking it afterwards. Args name the metric, horizon, and grain, e.g. /forecast monthly MRR for the next two quarters by segment.
Read analytics.md Part 6 first. For a predictive model about entities (which customers will churn) use /build-model; this skill is for a series over time.
Procedure
- Frame the use. Who will act on this forecast, on what decision, and how wrong can it be before the decision changes? That sets the horizon, the grain, and how much effort the intervals deserve. Ask whether they want a forecast (what we expect), and keep it separate from the target and the plan — if a number is going to be adjusted to match a goal, it is not a forecast and should not be labelled one.
- Define the series exactly. The metric from
knowledge/metrics-catalog.md, its grain, its period boundaries (knowledge/business-context.mdfiscal calendar), and whether the latest period is complete. Forecasting from a partial period is the most common silent error. - Check the history with
data-quality-engineer's lens: gaps, duplicates, definition changes, tracking migrations, and outages. Mark every known event from the change ledger on a plot of the series. Decide, and log, how each break is handled: truncate the history, model the break, or adjust the affected periods. Never interpolate over a data outage silently. - Look before modelling. Plot the series, its seasonal pattern, and its year-over-year view. Is there enough history — at least two full seasonal cycles to estimate seasonality at all? If not, say what the forecast can and cannot capture.
- Baselines. Compute naive, seasonal naive, and drift forecasts. These are the bar, and often the answer.
- Backtest — rolling origin, at the real horizon. Pick several cut-off dates across the last one or two years; for each, fit on data before it and forecast the horizon the stakeholder needs. Score WAPE (or MAE), and bias, per horizon step. Record the table. No random splits, no features unknown at forecast time.
- Candidate models, simplest first. Exponential smoothing / ETS, a seasonal ARIMA-class model, or a regression on trend, seasonality, and known drivers. Where the environment has one available, a pretrained time-series model may be added as another candidate — it goes through the same backtest. Add complexity only where the backtest says it pays. A model that does not beat seasonal naive in the backtest is not used; report the baseline as the forecast and say why.
- Intervals — and check them. Produce an 80% interval (and 95% if the decision needs it). Then measure empirical coverage in the backtest: the share of actuals that fell inside. If coverage is short of nominal, replace the model's interval with a conformal one built from the backtest's own error quantiles at each horizon step. Report the coverage achieved.
- Reconcile hierarchies. If the forecast is by segment or region, the parts must sum to the total. Forecast where the signal is best and reconcile, then check each level's backtest error still holds.
- Sanity-check against the business. Does the forecast imply a growth rate, a market share, or a capacity that is implausible? Does it contradict a known committed change (a price rise next month, a contract ending)? Known future events that the history cannot contain are added as explicit, labelled adjustments — never folded in invisibly.
- Write
models/<name>/FORECAST.md(and add a row tomodels/README.md):- Headline: the central forecast and its range, in one sentence a planner can use
- Forecast table: period, point, 80% low/high — beside, and separate from, any target
- What drives it and what it assumes continues
- Backtest results vs seasonal naive: error and bias by horizon, interval coverage
- Known risks not in the model, and explicit adjustments made
- How wrong it has been before — the plain-English version of the backtest ("at three months out, typically within ±6%")
- Retraining cadence, owner, and the provenance footer (
standards/reporting-standards.md)
- Ship the scoring path. A reproducible script: fresh history in, forecast table out (
models/<name>/forecast.py), seeds set, versions pinned. Schedule via {{ORCHESTRATOR or "documented manual run"}}. - Monitor. Hand
performance-monitorthe query that compares each realised period to its forecast, and the trigger: error outside the backtest range, or bias in one direction, for three consecutive periods means re-fit. Realised error goes on the monthly scorecard beside the metric.
Reporting rules
- Never a point forecast without a range. Never a range that stays the same width as the horizon grows.
- Round to the precision the uncertainty supports. A forecast of "$4,318,207" with a ±8% range is "$4.3M".
- State the as-of date of the data on every chart and table.
- When the forecast changes materially from the last one, say what moved it — new data, a changed assumption, or a model change — and by how much each.
Environment
Baselines, a backtest loop, and conformal intervals need only SQL plus a short Python script on the standard library. Statistical models need {{ML_ENVIRONMENT}} with statsmodels or similar; install on demand and pin versions in models/<name>/requirements.txt.
Created by Colin Beck — https://www.linkedin.com/in/beckcolin/
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

