Training models

RINK uses a Gradient Boosting Regressor with engineered lag and rolling-window features. Training happens automatically on upload, and you can retrain at any time from the Model card.

What a "trained model" means

For each upload (or each column switch), RINK fits a model on engineered features derived from the value column.

Features

For a target series y, the following features are constructed for each row:

Feature	Definition
lag1	y from one period earlier
lag2	y from two periods earlier
lag3	three periods earlier
lag5	five periods earlier
lag7	seven periods earlier
rmean3	mean of the previous 3 periods (excludes current)
rmean7	mean of the previous 7 periods

Rows with any NaN in the feature set (the first 7 rows by definition) are dropped before training. So a 100-row CSV produces ~93 training rows.

The model

GradientBoostingRegressor(
    n_estimators=200,
    learning_rate=0.05,
    max_depth=3,
    random_state=42,
)

These hyperparameters are a sensible default for univariate time series in the 100–10,000 row range. They're hard-coded in ml_api/main.py; tune them there if you have a specific benchmark in mind.

Validation split

The feature matrix is split 80/20 chronologically. The first 80% trains the model; the last 20% is used to compute RMSE and MAE.

This is a causal split — no future data leaks into training, which matches how you'd actually deploy a forecast.

Reading the metrics

After training, the RMSE and MAE KPI cards update:

• RMSE (root mean-squared error) — penalises large errors more heavily; useful when occasional big misses matter.
• MAE (mean absolute error) — average error magnitude; easier to reason about as "off by X on average".

Both are in the same units as your value column. So if your column is prices in dollars, MAE of 0.18 means "off by 18 cents on average across the validation set".

Retraining

Click Re-train Model in the Model card to fit a fresh model on the current dataset. This is useful when:

• You've switched the target column.
• You've replaced the CSV but the column is the same.
• You want to start from a clean random seed (the seed is fixed, but re-training clears any in-memory state).

Training time scales linearly with rows. For a typical 1,000-row CSV on Render Starter, training takes well under a second.

Persistence

After successful training, three files are written to your user directory:

/var/data/users/<your_uuid>/
├── uploaded.csv     # the original CSV, preserved as-is
├── model.joblib     # the fitted GradientBoostingRegressor
└── meta.joblib      # { column, date_column, frequency, days_per_step }

The model loads from disk on every prediction request, so you can sign out and back in (without auto-cleanup, e.g. by upgrading the cleanup-on-logout setting) and your model would still be available.

⚠️ By default, sign-out wipes these files. If you want them to persist across sessions, see the implementation note in the signOut wrapper.

When training fails

Symptom	Cause
Need at least 30 numeric rows	Fewer than 30 valid (non-NaN) values in the target column
CSV has no numeric columns	All columns are strings/empty
Failed to read CSV	File is corrupt, has unescaped quotes, or uses an unusual delimiter
Need at least 7 historical values	Tried to predict with too short an input
Training silently picks the wrong column	Use the column picker to override; see Switching columns