How to train a machine-learning model without writing code

One table, and a clear question. The table can be a CSV or an Excel file; each row is one thing you observed and each column is something you know about it. The question has to be answerable by one of those columns.

• Predicting a number — price, demand, yield, rate. This is regression.
• Predicting a category — will they churn, is this fraud, which grade. Classification.
• Predicting the future of a series — next month’s sales from this month’s. Forecasting, and a different tool.

How many rows is the question everyone asks. A few hundred is enough to learn something simple; a few thousand is comfortable. Below about a hundred, be sceptical of any result — there is not enough data for the held-out test to mean much.

A no-code tool runs a pipeline. Knowing what each stage does is the difference between trusting the output and hoping.

• Profile — how many rows, which columns are numbers, where the gaps are. Read this. Most bad results are visible here first.
• Split into train and test — and this comes early, before cleaning. Everything after it must be decided on the training half only, or the test is no longer a test.
• Handle missing values — drop, or fill. Which method to use.
• Encode text columns — models take numbers. A city name becomes a number or a set of yes/no columns.
• Scale — so a column in thousands does not drown one in single digits.
• Select features — keep the columns that carry signal. How to choose.
• Train several models and compare — never one. Different algorithms suit different data and you cannot tell which in advance.
• Evaluate on the test half — the one-time check on rows the model never saw.

The commonest mistake in no-code ML is not a wrong setting. It is cleaning the data before splitting it. If you fill missing values using the average of the whole table, the test rows helped compute that average — so they have influenced their own preparation, and the score you get at the end is optimistic by an amount nobody can measure.

Split first. Then fit every transformation on the training half and apply it to the test half. A tool that lets you clean before splitting will quietly flatter every result it gives you.

You will get a table of models and scores. The score alone means nothing — what matters is the gap to a baseline that ignores your data entirely. Check that first, every time, before deciding you have something.

Then check the gap between the training score and the test score. A small gap is healthy. A large one means the model memorised the training rows and will not generalise — fewer columns or a simpler model usually fixes it.