Financial intelligence

Market and alternative data pipelines

Most errors in quantitative finance are data errors. Prices need corporate-action adjustment for splits, consolidations and dividends. Backtests need delisted securities, or survivorship bias flatters every strategy. Fundamentals must be point-in-time, as reported on the day, not as later restated, or the model learns from information it could not have had. Timestamps must carry their time zone and the AEST to AEDT transition.

We store financial data bitemporally, recording both when a fact was true and when we learned it, so any historical run can be reproduced exactly. Alternative data such as satellite imagery of car parks and stockpiles, vessel AIS movements, web traffic, job postings and card-spend panels is ingested with its licence terms attached to the dataset, because redistribution and use rights differ by vendor.

NLP over filings and announcements

Listed entities must disclose price-sensitive information immediately under ASX Listing Rule 3.1, the continuous disclosure rule (ASX Listing Rules). That makes the announcements platform one of the richest event streams in the market. Our pipelines take announcements from a licensed feed or your existing data vendor and:

• Parse PDFs with layout-aware models, so tables, footnotes and multi-column text survive extraction.
• Classify announcement type and extract structured events: guidance changes, capital raisings, director interest notices (Appendix 3Y), substantial holder notices, contract wins and trading halts.
• Run finance-tuned language models for extraction, with outputs validated against a schema and every extracted value citing its page and line.
• Measure latency from release to structured signal, because an extraction that arrives late has a different value.

Entity extraction and resolution

A model that cannot tell which company a sentence is about is worse than no model. We combine named-entity recognition with entity linking to stable identifiers: ASX codes, ISINs, Australian Business Numbers from ABN Lookup and Legal Entity Identifiers from GLEIF, including GLEIF's parent relationship data for corporate hierarchies. Name changes, mergers and subsidiaries are handled as dated edges in a knowledge graph of ownership and directorships, which also feeds our money laundering intelligence work.

Credit and counterparty risk models

Credit models estimate probability of default (PD), loss given default (LGD) and exposure at default (EAD). For authorised deposit-taking institutions using the internal ratings-based approach, APRA's APS 113 sets the requirements those estimates must meet. We build:

• Scorecards. Weight-of-evidence binning and logistic regression remain the standard where every coefficient has to be defended.
• Constrained machine learning. Gradient-boosted models with monotonic constraints, so a worse arrears history can never improve a score, benchmarked against the scorecard as a challenger.
• Early-warning signals. Changes in transaction behaviour, payment timing, utilisation and adverse news that move a borrower onto a watchlist months before a default.
• Counterparty exposure. Monte Carlo simulation of exposure profiles per netting set, giving expected and potential future exposure, with wrong-way risk tested where exposure and default probability rise together.

Our founder, Kunal Miind, consults to financial institutions on AI, including financial modelling and conversational analysis, and has built fraud detection and machine learning models for clients including Philip Morris, Queensland Health and Australian Government agencies.

Conversational analysis

Calls, chats and emails hold signals that never reach a ledger: complaints, hardship, possible misconduct, mis-selling and the reasons customers leave. We transcribe recorded calls with speech-to-text and speaker diarisation, then classify and summarise conversations so compliance and customer teams review the ones that matter instead of a random sample. Personal information is redacted before analysis, models run in your environment where required, and every flag links back to the recording and the passage that triggered it.

Forecasting with error ranges

We forecast cash flow, deposits, collections, claims and revenue as distributions, not single lines. Hierarchical forecasts are reconciled so branch, region and total add up. Prediction intervals come from quantile models and conformal prediction, which gives calibrated coverage without assuming normal errors. Every model is backtested on rolling origins that match how it will be used each week or month, and scenario runs take macroeconomic drivers such as the cash rate path as explicit inputs.

Anomaly detection

Unsupervised and semi-supervised detectors surface what rules do not describe: journal entries posted by unusual users, at unusual times or to unusual account combinations; expense claims that break a person's own pattern; order-to-trade ratios that warrant an internal surveillance review; and data-quality breaks such as a vendor feed that silently changes units. Every flag lists the features that drove it.

Explainability and model risk

Models that make or support decisions need governance an auditor can test: a model inventory, independent validation with challenger models, documented assumptions and limits, and ongoing monitoring of discrimination (AUC, Gini) and population stability. Individual decisions come with SHAP attributions, interpreted with care where features are correlated, and counterfactual explanations for adverse outcomes. For APRA-regulated entities, CPS 230 (operational risk management, including service providers) and CPS 234 (information security) shape how platforms and models are hosted, monitored and supplied, and we document controls against them.

Data lineage and audit

Pipelines load raw data unchanged, transform it in versioned and tested steps, check data contracts at every boundary and reconcile totals back to the ledger or source system on each run, stopping loudly on a mismatch. Column-level lineage records which source rows produced each reported number, and logs are append-only. Any past report can be regenerated from the data and code that produced it.

Questions