Lecture 10 - Advanced Auditing

Why is innovation needed in auditing?

Large datasets, digitalization, better quality, regulatory pressure.

What technologies drive modern audits?

AI, deep learning, data analytics, continuous auditing.

What is an autoencoder?

Unsupervised neural network that reconstructs inputs and detects anomalies.

What does reconstruction error indicate?

How unusual a journal entry is.

What is the Accounting Anomaly Score (AS)?

Combination of reconstruction error and attribute probability.

Key autoencoder limitations?

Replicates errors in training data, difficult to interpret, can miss camouflaged fraud.

What is Isolation Forest?

Tree-based model isolating anomalies with few splits.

Why is Isolation Forest fast?

No training, random splits, scales to millions of entries.

How to improve stability in Isolation Forest?

Use many trees (≈100) or fix random seed.

Autoencoder vs Isolation Forest (main difference)?

Autoencoder learns patterns; IF isolates anomalies via random splits.

What is continuous auditing?

Automated, high-frequency, full-population testing.

Key steps of continuous auditing?

Define, gather, analyze, detect, handle, review.

Why is continuous auditing different from data analytics?

High frequency, automation, exception-focused, integrated workflows.

Examples of continuous controls?

3-way match, duplicate invoices, SoD, vendor changes.

Purpose of Did Do testing?

Identify what actually happened (exceptions).

Purpose of Can Do testing?

Verify what the system enforces (application controls).

Why does AI not replace auditors?

Human judgment needed for interpretation, fraud detection, model risks.

What skills must auditors develop?

AI literacy, ITGC knowledge, analytics interpretation, skepticism.