When TensorFlow reports 0.99 accuracy but real predictions are poor, the model is usually optimizing the wrong thing for your problem setup. Accuracy alone can hide major failure modes.

Why this happens

1. Class imbalance

If 99% of samples are class A, a model that always predicts class A gets 99% accuracy while being useless for class B.

2. Data leakage

If training features include future information or label-derived signals, validation accuracy can look excellent but collapse in production.

3. Overfitting

The model memorizes training patterns and does not generalize to unseen examples.

4. Train/serve mismatch

Different preprocessing in training vs inference (normalization, tokenization, label encoding, image resize) causes bad live predictions.

5. Wrong threshold

For probabilistic classifiers, a default threshold of 0.5 may be suboptimal. Accuracy can look high while precision/recall for the target class is poor.

What to check first

1. Inspect class distribution in train, validation, and test splits.
2. Plot a confusion matrix, not just a single scalar metric.
3. Report precision, recall, F1, and PR-AUC/ROC-AUC.
4. Verify preprocessing parity between training and inference pipelines.
5. Evaluate on a truly held-out dataset from the same production distribution.

Metric formulas (plain text)

• Precision = TP / (TP + FP)
• Recall = TP / (TP + FN)
• F1 = 2 * (Precision * Recall) / (Precision + Recall)

TensorFlow-focused fixes

• Use class weights: model.fit(..., class_weight=...)
• Use stratified splits when possible.
• Add regularization (dropout, weight decay) and early stopping.
• Calibrate probability thresholds on validation data.
• Use data augmentation only when it preserves label meaning.

Minimal diagnostic workflow

1. Train baseline model and save per-class metrics.
2. Generate confusion matrix on validation set.
3. Tune decision threshold for target business metric.
4. Re-run evaluation on held-out test set.
5. Sanity-check with manual examples from production.

Summary Table