Research Methodology

This page describes the general research methodology of AlgoGreek Research. It does not disclose proprietary source code, confidential parameters, or live model weights.

Quantitative research

Research outputs may be prepared using quantitative, algorithmic, mathematical, statistical and data-driven methodologies, used independently or together with traditional fundamental and technical analysis. The process is systematic and rules-based, designed to minimize discretionary bias. Coverage may include equities, ETFs, derivatives and indices listed on recognised stock exchanges.

Statistical analysis

Statistical analysis is used to describe distributions of returns, volatility, breadth, factor exposures and related market observations drawn from publicly available information and known observations. Statistical description is not a prediction of future prices.

Algorithmic research

Algorithmic components encode rules for ranking, simulation, position-sizing logic and risk controls. PlutusQuant, as described on the house page, evaluates names in the investable universe using a systematic simulation that incorporates volatility, position sizing and risk controls; names with the strongest simulated performance are ranked and the top ten are selected for the model portfolio. A separate global Regime model classifies risk-on, neutral or defensive climate and sets risk for the books. Model portfolios hold those books; they are illustrative research constructs.

Model development

Absolute controls are maintained over the lifecycle of model development, code testing, historical verification, systematic implementation, subsequent modifications and periodic reviews. All material modifications, mathematical parameter adjustments or structural rule modifications to active research models are documented and logged, including model identifier, version, release date, methodology version, research universe, description of changes and reason for change.

Model validation

Before a model version is used to produce research for clients, it is subject to validation appropriate to its design: theoretical review of the rule set, checks against data quality, and comparison of in-sample behaviour with the model’s stated objective. Validation does not guarantee out-of-sample results.

Historical testing and backtesting

The desk distinguishes three different records. They must not be mixed.

• Backtest. A historical simulation of how a model version would have behaved on past data. Backtested results are not actual investor performance, are not a track record of client accounts, and must never be labelled as such. They are sensitive to assumptions on fills, costs and look-ahead.
• Live model. The model version that is currently in production for the day’s book. Its identifier and version are stored with each published research output.
• Actual research recommendation history. The immutable record of what was published to clients: instrument, rating, target, stop, timestamp, model version and disclosure version. Corrections create a new revision; they do not overwrite the original.

Backtested results are not presented on this website as marketing performance (for example as a rupee path or a CAGR) unless and until such disclosure is expressly enabled under the applicable SEBI / PaRRVA framework. The data model allows proper performance reporting in the future without rewriting history.

Out-of-sample testing

Where used, out-of-sample tests evaluate a frozen rule set on data that was not used to specify the rules. Passing an out-of-sample test is not a forecast of live results.

Walk-forward validation

Where used, walk-forward validation rolls the estimation window forward in time and re-evaluates. It is a research-control technique, not a live trading account.

Data quality

Inputs are taken from market data and company information believed to be reliable. AlgoGreek Research does not warrant completeness or absence of error in third-party data. Corporate actions, missing prints and delayed feeds can affect ranks.

Transaction costs and slippage

Simulations may include stylised costs. Actual investor costs, taxes, impact and slippage will differ. Model portfolios are not executable orders.

Model limitations

Models can fail when regimes change, when liquidity disappears, or when the data-generating process shifts. Rankings can cluster. A top-ten book is concentrated. Derivatives, if covered, have additional path-dependent risk. AI / ML tools, where used, are subject to the limitations on the AI Disclosure page. Quantitative outputs remain fully subject to disclosure, conflict management, personal-trading bars and compliance requirements applicable to research reports.

Periodic model review

Active models are reviewed periodically. Retirement of a version is recorded. Historical research reports continue to reference the model version that existed when the report was published. Updating a model does not modify historical reports. Research reports, recommendations and the rationale for recommendations are retained for a minimum of five years.