Mean reversion for regular savings: the method and how it is tested

Summary

Most Indian savers invest the same amount every month, whatever the market is doing. This paper describes a different rule. A model estimates how far a fund or index is trading from its own long-run trend, and the rule invests more when it is meaningfully below that trend and less when it is above. Everything stays within a budget cap the investor sets, and the rule runs alongside a regular plan, never instead of it.

The paper explains the idea, the model, the rule, and how the rule is tested. It does not report results. SEBI's advertising rules restrict past-performance figures for Research Analysts to those verified by the Past Risk and Return Verification Agency (PaRRVA), so this paper contains none. It is meant to let a reader judge whether the method and the test are sound before seeing any result.

1. The question

Can a saver who already invests a fixed amount every month do something useful with an extra budget by timing how much goes in, based on how cheap the asset is against its own history?

And just as important: how would anyone know? An answer is only worth something if the test behind it is fair. Most of this paper is about the test.

2. Mean reversion, briefly

Mean reversion is the statistical tendency of an asset's price to drift back toward its long-term trend after moving away from it.

A mean-reversion method does not try to predict next week's price. It measures the size of the gap between the current price and the trend, and acts on that gap. A large gap below the trend is treated as a better moment to add money than a gap above it.

Mean reversion is a long-run tendency, not a certainty. A trend can shift. An asset can stay cheap for a long time or get cheaper. A method built on mean reversion will at times invest early into a fall that keeps going. The budget cap in section 3 exists for exactly this reason.

3. The method

The method has three parts.

3.1 The trend and the gap

For each fund or index, the model looks only at that asset's own published history of prices or net asset values (NAVs). From that history it estimates a long-run trend, then measures how far the latest price sits from it. Each asset is compared with itself, not with other assets.

3.2 The model

The gap is estimated by an ensemble of machine-learning models, which are several models whose estimates are combined, so that no single model's quirks decide the outcome. The models are built and run by the analyst. Their output is one number per asset per period: how cheap or dear the asset is against its own history.

The exact inputs and model settings are not published, because they are the analyst's own research. What is published is what the model does, what it is allowed to see, and how it is tested.

3.3 The rule

The investor sets a fixed maximum for each period, for example a weekly cap. Each period, the rule turns the model's estimate into an amount between zero and that cap:

• when the asset is meaningfully below its trend, the amount moves toward the cap;
• when it is near or above its trend, the amount moves toward zero;
• the amount never exceeds the cap.

The rule runs alongside the investor's regular fixed plan, as extra money on top. It does not replace a regular plan, and it never borrows ahead or invests more than the cap.

4. How the method is tested

A test of an investing rule can easily flatter it. The design below is meant to make that hard.

4.1 Data

Tests use published NAV histories of mutual fund schemes and published index levels, with history from 2009 onwards.

The test universe covers a broad set of equity schemes across fund houses and styles, and the main broad-market indices, for example the Nifty 50, Nifty Next 50, Nifty Midcap 150 and Nifty Smallcap 250. The securities quoted are for illustration only and are not recommendatory.

4.2 The comparison

The rule is always compared with a fixed recurring plan in the same asset over the same dates. Same asset means no fund-picking skill can creep in. Same dates means both face the same market.

Because the rule invests different amounts each period, a simple percentage change would mislead. Both are measured by XIRR, the money-weighted annual return that accounts for when each rupee went in. This is the same measure a saver sees on a mutual fund statement.

4.3 Every start date, not a chosen few

Any rule can be made to look good by picking a convenient start date. So the test starts the rule and the fixed plan in every month of the test period, and runs each for fixed horizons of one to five years. Each start month gives one result per horizon.

The results are read as a distribution: the typical outcome, the middle half of outcomes, and the full range from worst to best. The worst cases matter as much as the typical one.

4.4 Data the model has not seen

A model is only tested on data it was not built on.

• Fitting. The model is fitted on an earlier part of the history.
• Out-of-sample test. It is then run, unchanged, on later years it never saw. Out-of-sample testing runs from 2021 onwards.
• Live use. Since November 2025 the model has run on new market data as it arrives. Before registration on 1 June 2026 this was internal testing only, and no recommendations were given.

4.5 Different markets

The test period covers several distinct market conditions, including the recovery after the 2008 global financial crisis, the 2013 taper tantrum, the 2018 mid- and small-cap fall, the 2020 COVID crash and recovery, and the 2022 rate-shock correction. A rule that only works in one kind of market is not a rule.

4.6 What the test cannot tell you

• The past is not the future. A test shows how a rule behaved on history. Markets change, and a pattern that held may stop holding.
• Long horizons. Over very long periods, the market's own drift dominates, and the timing of extra contributions matters less.
• Costs and taxes. How the test treats exit loads, stamp duty and taxes is stated alongside any published result.
• Behaviour. A test assumes the saver follows the rule every period. Real savers sometimes stop, especially in a fall, which is exactly when the rule invests most.
• Model risk. The model can be wrong. The cap limits how much money a wrong estimate can put to work in one period, but it does not remove the risk.

5. How results will be published

Results are not in this paper. If they are published later, they will appear only after verification by PaRRVA, as SEBI requires, and will show the full distribution across start months, worst cases included, with the data sources.

6. Use of AI in this research

SEBI requires a Research Analyst to disclose the extent of AI use in research services (MC ¶1.7).

• Machine-learning models. The gap estimate in section 3.2 comes from machine-learning models built and run by the analyst. Their output decides the amount the rule invests each period.
• Responsibility. The analyst is solely responsible for these models, their output, any research based on it, and the security of any client data used with them.

7. Where the method is applied

The method is applied in Rezz Invest, the analyst's own research application for regular savers, offered under the same SEBI registration. See the Disclosures page for related disclosures.

8. Disclosures

• The Research Analyst is Tushar Gerg, SEBI-registered Research Analyst, Reg. No. INH000027654, registered from 01 June 2026. Registered office: A-302, Plumeria Garden Estate, Omicron Sector 3, Noida, Uttar Pradesh, 201308. Phone: +91 9997976746. Contact and complaints: verifiedresearch.in/regulatory/complaints/.
• Compliance officer, grievance officer and principal officer: Tushar Gerg · tushar0708@gmail.com · +91 9997976746.
• The analyst is registered as a part-time Research Analyst. The other activity is analytics consulting, which is unrelated to securities and is kept at arm's length from this research.
• The Research Analyst or any of its officer/employee does not trade in securities which are subject matter of recommendation.
• Research Analyst or its employee or its associates have not received any compensation or other benefits from the Subject Company or 3rd party in connection with the research report/ recommendation.
• This paper does not recommend any security. It describes a method.
• This paper is research, not investment advice. It does not consider any reader's circumstances.
• The Research Analyst cannot execute trades on a client's behalf.
• Full disclosures: verifiedresearch.in/regulatory/disclosures/

Sources

1. SEBI, Master Circular for Research Analysts, 06 February 2026. https://www.sebi.gov.in/sebi_data/attachdocs/feb-2026/1770375507051.pdf

Tushar Gerg, SEBI-registered Research Analyst, Reg. No. INH000027654 · Complaints: verifiedresearch.in/regulatory/complaints/ · SEBI SCORES · SMART ODR