YourCloudGroup Model Manager is a browser-based software-as-a-service platform that builds, evaluates, certifies and schedules predictive models — classification and regression — from a customer's own structured data. It is made by YourCloudGroup.

It is designed for business users who want predictions from their own spreadsheet data without writing code or employing data scientists. Users provide structured data, such as a CSV file or a Google Sheets link, and choose the column they want to predict. YourCloudGroup Model Manager freezes a training/blind split of that data, screens every column for information that would not exist at prediction time, trains twelve candidate models, scores each on blind rows it never saw, and picks the winner on balanced measures rather than raw accuracy. Each model then receives a certificate made of fourteen independent, evidence-based claims. It does not generate text: it is a predictive-modeling tool, not a large language model.

Who is YourCloudGroup Model Manager for?

YourCloudGroup Model Manager is for small and medium-sized businesses, individuals, and teams who have data in spreadsheets and want predictions from it — people for whom writing Python, maintaining notebooks or running a machine-learning infrastructure is not practical. An API is also available for technical teams who want to call models from their own systems.

What problem does it solve?

Building a predictive model is not the hard part. Knowing whether its score can be trusted is. A single accuracy figure is a claim, and most claims in machine learning are never checked: a column filled in only after the outcome is known, test rows reused until the model is tuned on them, or a retrained model that quietly comes out different will all inflate a score without showing up in it.

YourCloudGroup Model Manager combines modelling and that checking into one point-and-click pipeline in the browser:

Split data — divide one labelled dataset into a training set and a sealed blind set, and freeze the experiment record.
Build model — train candidate models on the training rows only.
Tournament — compare twelve model variants on the same blind rows and pick the winner on balanced measures.
Predict — run the winning model on new rows.
Evaluate — measure results against known outcomes, including outcomes that arrive weeks later.
Schedule — run predictions and training automatically.

What kinds of data does it use?

YourCloudGroup Model Manager works with structured, tabular data: rows and columns, where each row is one case and one column is the outcome to predict. Data can be loaded by uploading a file such as a CSV, or from a Google Sheets link, an S3 presigned URL, a Dropbox share link or a GitHub raw link. Scheduled runs can fetch fresh data from Google Sheets or S3 on each run.

It is not designed for images, audio or free-form text documents.

What kinds of models does it build?

YourCloudGroup Model Manager builds classification models (predicting a category, such as "will renew" or "will not renew") and regression models (predicting a number, such as an order value). It detects which task applies from the target column.

For each task it runs a model tournament: twelve model variants across six architectures — neural networks of three shapes, random forest, gradient boosting and logistic regression. All twelve are trained on the same rows and scored against the same blind set. The winner can then be rebuilt on all of the training data, and the rebuild is scored on the same held-out rows, so the two figures are directly comparable. Existing models can be extended with new data through continuing training.

How are models evaluated?

Every model is evaluated against a blind set: rows held back when the data is split, whose answers stay sealed away from the model. The tournament winner is chosen on balanced accuracy and macro F1, not on raw accuracy, which rewards a model for predicting the majority class. The evaluation suite reports:

For regression: R-squared, mean absolute error and root mean squared error.
For classification: confusion matrices, balanced accuracy, macro F1 and per-class recall.
For both: confidence analysis and row-by-row results.

A metric is only reported where it means something. The blind set is checked against the model's declared target domain — the outcomes the model was built to predict. If a declared class is missing from the blind set, the evaluation is inadequate: accuracy stays visible as descriptive, while balanced accuracy, macro F1 and per-class recall read NOT MEASURABLE, with the raw figure shown beside them. Because a model records the outcomes it was fitted on, it can still be graded properly from its own artifact long after its original data has been cleared away.

A worked example of why balanced measures matter is our blind evaluation of a fraud classifier on 28,306 transactions.

How does YourCloudGroup Model Manager check a model's score is honest?

The governing rule is:

A certified model must never gain accuracy from information that would not exist when the prediction is actually requested.

The checks happen where the experiment is defined — when the data is split — not after the model is built. The split is frozen: a record of what the target means, which rows went where, and which columns may be used is fixed and cannot be edited afterwards. Every later step reads that record. The split decides what the data means; the model build only does the modelling.

Is any column the answer in disguise? Each candidate column is examined on the training rows only — never the blind rows — for direct leaks that reproduce the outcome almost exactly, for leaks that live in a group of low-cardinality columns that together rebuild the outcome, for identifiers such as customer IDs and row numbers, and for post-event fields that are filled in only once the outcome is known. A very strong but legitimate predictor is sent to a person for review rather than excluded for being good. A column's name is evidence, never a verdict.
When does the prediction happen? The experiment declares the prediction moment: a timestamp column, a time built from several columns, a named business event such as "at booking", or the moment each observation arrives. For data without dates, that last option requires an explicit declaration; it is never assumed, and the declaration becomes part of the frozen record.
Did every outcome have a fair chance to become known? The platform measures how long outcomes take to appear across the training population and reports whether every record had a fair chance to reach its outcome — see outcome observability.
Can this blind set support a claim at all? Before the answers are opened, the blind set is checked for size, open records and missing outcome classes — evaluation adequacy. Reusing an evaluation population seen before is disclosed provenance that qualifies the claim; contamination, where earlier evaluation rows leaked into training, is an integrity failure that blocks certification.

What does "certification" mean here?

In YourCloudGroup Model Manager, a certificate is not one stamp. It is fourteen independent claims, each with its own state and a stable reason code an auditor can trace, grouped into three kinds of assurance:

Assurance	Claims
Model assurance — is this the model we say it is?	Identity · Lineage · Training-data boundary · Target domain · Feature schema · Build reproducibility · Inference repeatability
Performance assurance — did it earn its score honestly?	Observed predictive performance · Evaluation reproducibility · Blind integrity
Deployment assurance — can it be trusted in its real setting?	Prediction-time availability · Outcome observability · Evaluation feasibility

Each claim is one of CERTIFIED, NOT ESTABLISHED, NOT ASSESSED, NOT APPLICABLE or FAILED. The claims are reported side by side and never merged into a single verdict. See certification claim states.

The model's status comes from the model-assurance claims. A model whose prediction moment was never declared can still be a certified model; it is not a certified deployment, and the certificate says so in a separate section.
The performance claim needs three things: the metric was actually measured, the evaluation was adequate, and training and evaluation were isolated. A strong number from an inadequate test does not qualify.
"Not assessed" is an answer. When the data cannot support a claim, the claim reads NOT ASSESSED and the certificate explains why and what would change it, rather than filling the gap with an assumption.
An issued certificate is never re-judged. When the methodology improves, older models get a clearly labelled current assessment rebuilt from their frozen evidence, next to the certificate originally issued. Every claim is worked out again from the frozen records rather than from stored text, so a certificate cannot drift from what actually happened.

YourCloudGroup also runs a hands-on lab, about half a day long, that certifies people as AI model builders. That is a separate offering from model certification.

What does "repeatable" mean here?

A model earns the REPEATABLE designation only when the inference-repeatability check actually ran and passed: the frozen model is asked the same questions twice and the outputs are compared. Repeatability is verified, never assumed because an algorithm is supposed to be deterministic. Build reproducibility — whether rebuilding the model produces the same model — is reported as its own, separate claim.

Operationally, trained models can also be run again on new data on a schedule — hourly, daily, weekly or monthly — on the customer's dedicated server, which has daily backups.

What does "audit" mean here?

A model is designated AUDITABLE when the evidence behind its certificate is complete but the repeatability check never ran. More generally, every claim carries a stable reason code, and the frozen experiment record, the evaluation results, the confusion matrix and row-by-row results let a user or an auditor check how a score was produced rather than relying on a single summary number. A failure on any certification gate withholds both designations.

Certification here is YourCloudGroup's own evidence-based methodology. It is not a third-party audit or a regulatory compliance report.

What is and is not being certified?

Is certified: specific, stated properties of one model — each of the fourteen claims that reads CERTIFIED — for example that the artifact is the one described, that it was fitted only on its declared training boundary, and that its performance was measured on an adequate, isolated blind set. Separately, a person who completes the hands-on lab is certified as an AI model builder.
Is not certified: future performance. Markets shift, customers change, and a model that was honest yesterday can go stale. Certification tells you the score was earned honestly, under conditions you can inspect; it does not tell you the world will stay the same. Claims that read NOT ASSESSED, NOT ESTABLISHED or NOT APPLICABLE are exactly that — not certified — and deployment is not certified unless the deployment-assurance claims are.

How is it different from a large language model?

A large language model generates text from a prompt, using patterns learned from very large general-purpose text collections. YourCloudGroup Model Manager learns from a customer's own historical rows with known outcomes and outputs a category or a number for each new row. Its results can be measured directly against blind data with known answers, and certified claim by claim. For a fuller comparison, see predictive AI vs. generative AI.

What are its limitations?

Structured data only. It is built for tabular data, not images, audio or free-form documents.
Not a text generator. It does not write, summarise or converse.
Quality depends on the data. A model can only learn patterns that exist in the columns provided. Missing history or too few examples of a rare outcome limit results, and leakage screening can only exclude what the data reveals.
Some claims need declarations. Deployment claims such as prediction-time availability depend on the experiment declaring when predictions happen; without that, they read NOT ASSESSED.
Certification is not a guarantee of future performance. It certifies how a score was earned on the stated blind set, not how the model will perform if conditions change.
Business judgement is still required. The platform reports how often a model is right, where it is wrong and which claims are established; deciding whether that is good enough for a given decision is up to the user.

Where can I learn more?

The official website: yourcloudgroup.com, including sign-up and current pricing.
A free live webinar every Thursday at 1:00 PM Pacific, listed on yourcloudgroup.com.
On this blog: can you build an AI model from a CSV file? and what held-out model evaluation is.
About YourCloudGroup and our editorial standards.