What causes "model is not fitted yet!" in microsoft/qlib?

Run model.fit(dataset, evals_result) to completion before predict; verify it logs 'best score: ...' (the signal fit finished).

microsoft/qlib · error · ValueError

model is not fitted yet!

Error message

model is not fitted yet!

What it means

The TS (time-series) LOCALTransformer variant's predict() checks self.fitted before preparing the 'test' segment. self.fitted becomes True only at the end of a successful fit(); any predict before that raises ValueError('model is not fitted yet!') without touching the dataset.

Source

Thrown at qlib/contrib/model/pytorch_localformer_ts.py:205

                stop_steps = 0
                best_epoch = step
                best_param = copy.deepcopy(self.model.state_dict())
            else:
                stop_steps += 1
                if stop_steps >= self.early_stop:
                    self.logger.info("early stop")
                    break

        self.logger.info("best score: %.6lf @ %d" % (best_score, best_epoch))
        self.model.load_state_dict(best_param)
        torch.save(best_param, save_path)

        if self.use_gpu:
            torch.cuda.empty_cache()

    def predict(self, dataset):
        if not self.fitted:
            raise ValueError("model is not fitted yet!")

        dl_test = dataset.prepare("test", col_set=["feature", "label"], data_key=DataHandlerLP.DK_I)
        dl_test.config(fillna_type="ffill+bfill")
        test_loader = DataLoader(dl_test, batch_size=self.batch_size, num_workers=self.n_jobs)
        self.model.eval()
        preds = []

        for data in test_loader:
            feature = data[:, :, 0:-1].to(self.device)

            with torch.no_grad():
                pred = self.model(feature.float()).detach().cpu().numpy()

            preds.append(pred)

        return pd.Series(np.concatenate(preds), index=dl_test.get_index())

View on GitHub (pinned to 79633dd950)

Solutions

Run model.fit(dataset, evals_result) to completion before predict; verify it logs 'best score: ...' (the signal fit finished).
Diagnose and fix any earlier fit() failure — fitted stays False until a clean finish.
To restore a trained model: torch-saved best_param exists at save_path; recreate the model, model.model.load_state_dict(torch.load(save_path)), set model.fitted = True, then predict.

Example fix

# before
model = LOCALTransformerModel(...)
preds = model.predict(dataset)  # ValueError: model is not fitted yet!

# after
model = LOCALTransformerModel(...)
model.fit(dataset, evals_result)
preds = model.predict(dataset)

Defensive patterns

Strategy: validation

Validate before calling

if not getattr(model, "fitted", False):
    raise RuntimeError("model not fitted; run fit() (or restore checkpoint + set fitted=True) first")
preds = model.predict(dataset)

Type guard

def is_fitted(model) -> bool:
    return bool(getattr(model, "fitted", False))

Try / catch

try:
    preds = model.predict(dataset)
except ValueError as e:
    if "not fitted" in str(e):
        model.fit(dataset, evals_result)
        preds = model.predict(dataset)
    else:
        raise

Prevention

Assert model.fitted before every predict in production code.
Treat checkpoint restore as two steps: load_state_dict AND fitted = True.
Abort workflows on fit failure so predict is never reached.

When it happens

Trigger: Calling model.predict(dataset) on a LOCALTransformerTS-style model where fit() was never run, failed partway (e.g. empty data, NaN loss, OOM), or was skipped in a scripted workflow.

Common situations: Workflow configs where the train task failed but downstream predict/backtest tasks still executed; re-loading an unfitted pickled model; interactive notebooks jumping to evaluation.

Related errors

AI-assisted analysis of microsoft/qlib@79633dd950 (2026-08-15). Data as JSON: /api/errors/852818946af9b6d4. Report an issue: GitHub.