How did Mean-variance (shrunk covariance) do after costs and tax?
Long-only maximum Sharpe ratio with the Ledoit-Wolf covariance and each asset's mean excess return over the last 60 months (at least 36 when less history exists); minimum variance if no asset has a positive expected excess return; monthly. From Apr 2005 to Oct 2026 it compounded at 9.3% a year after costs and tax (10.7% before tax), against 11.4% after tax for 60/40. Its worst fall was −41%, bottoming in Oct 2008 and recovered by May 2010.
Written for investors weighing this allocation against a plain 60/40, and advisers who want its tax and cost drag measured. Data to 5 Oct 2026.
How it works
One of 136 portfolios in the lab, in the family "Risk-based". The rule and its parameters were written down before any result was computed.
Historical research on index portfolios, not investment advice. Returns are after fund costs; 'after tax' applies Indian capital-gains rules by date for a 30% slab investor and sells everything at the end. Most factor, mid-cap and small-cap index history before each index's launch date is back-tested by NSE. One 21-year sample, domestic assets only.
- Construction
- Long-only maximum Sharpe ratio with the Ledoit-Wolf covariance and each asset's mean excess return over the last 60 months (at least 36 when less history exists); minimum variance if no asset has a positive expected excess return; monthly.
- Rebalancing
- A rule sets the weights at each month-end close; trades happen at the next session's close.
- From the literature
- Markowitz (1952); Ledoit and Wolf (2004); see DeMiguel et al. (2009) on its out-of-sample record.
| Asset | Target | Average held | Lowest | Highest | Latest (Oct 2026) | More |
|---|---|---|---|---|---|---|
| G-sec 5-year (NSE benchmark index) | – | 38% | 0% | 94% | 0% | |
| Nifty 500 | – | 32% | 0% | 100% | 23% | |
| Gold (domestic, rupees) | – | 29% | 0% | 77% | 77% |
Growth and falls
What one rupee became, and how far it fell on the way. 60/40 is drawn in grey for comparison. Both curves are after fund costs and before tax; the after-tax figures are in the costs chapter.
Growth of 1 rupee
Weekly, shown month by month, Apr 2005 to Oct 2026. After costs, before tax. Log scale.
How to read it. On a log scale equal slopes are equal rates of return. One rupee became 8.89× here and 11.21× in 60/40; a steeper line in one stretch says which did better then.
Fall from the previous peak
The lowest weekly reading of each month, per cent below the highest value reached so far.
How to read it. Zero means a new high. The deepest fall here was −41.3% (Oct 2008), against −36.5% for 60/40. Depth is half the story: the width of each dip is how long an investor waited to get back to even.
per cent
The five deepest falls
| Fall | Peak | Bottom | Back to peak | Days falling | Days below peak | More |
|---|---|---|---|---|---|---|
| −41.3% | 9 Jan 2008 | 24 Oct 2008 | 12 May 2010 | 289 | 854 | |
| −16.3% | 10 May 2006 | 14 Jun 2006 | 10 Nov 2006 | 35 | 184 | |
| −14.5% | 29 Jan 2026 | 23 Mar 2026 | not yet | 53 | 249 | |
| −12.4% | 29 Jan 2015 | 7 Sept 2015 | 3 Oct 2016 | 221 | 613 | |
| −9.4% | 26 Nov 2012 | 7 Aug 2013 | 7 Mar 2014 | 254 | 466 |
Year by year
Calendar-year returns, after costs and before tax, with 60/40 below for comparison. In 8 of 20 full years this portfolio did better. The market phases below are the lab's pre-registered splits.
| 2005* | 2006 | 2007 | 2008 | 2009 | 2010 | 2011 | 2012 | 2013 | 2014 | 2015 | 2016 | 2017 | 2018 | 2019 | 2020 | 2021 | 2022 | 2023 | 2024 | 2025 | 2026* | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Mean-variance (shrunk covariance) | +18 | +17 | +51 | −30 | +41 | +15 | +6 | +10 | +1 | +13 | −3 | +10 | +13 | +2 | +10 | +13 | +6 | +6 | +14 | +11 | +29 | +2 |
| 60/40 | +24 | +22 | +40 | −27 | +51 | +10 | −15 | +23 | +5 | +29 | +4 | +8 | +24 | +1 | +9 | +15 | +20 | +3 | +19 | +13 | +8 | −3 |
| Phase | Dates | A year | 60/40, a year | Worst fall | 60/40, worst fall | More |
|---|---|---|---|---|---|---|
| The boom, FY06 to FY08 | Apr 2005 to Mar 2008 | 20.4% | 20.9% | −21% | −22% | |
| The crisis, FY09 | Apr 2008 to Mar 2009 | −8.8% | −14.6% | −32% | −27% | |
| FY10 to FY14 | Apr 2009 to Mar 2014 | 12.7% | 14.4% | −9% | −17% | |
| FY15 to FY20 (to March 2020) | Apr 2014 to Mar 2020 | 6.8% | 7.9% | −12% | −23% | |
| April 2020 on | Apr 2020 to Oct 2026 | 12.1% | 14.6% | −14% | −11% |
What costs and tax took
The same Apr 2005 to Oct 2026 run, peeled back one layer at a time. Fund costs are those of the cheapest widely available index fund or ETF of each era; tax applies Indian capital-gains rules as they stood on each sale date, and everything is sold at the end so deferred tax is counted. Turnover was 0.80 times the portfolio a year.
| Layer | Mean-variance (shrunk covariance) | Cost of this layer (pts) | 60/40 | More |
|---|---|---|---|---|
| Index return, before any cost | 11.57% | – | 12.48% | |
| After fund costs and trading (before tax) | 10.68% | 0.89 | 11.89% | |
| After tax, 30% slab | 9.32% | 1.36 | 11.41% | |
| After tax and inflation | 2.70% | 6.62 | 4.67% |
| Tax case | Mean-variance (shrunk covariance) | 60/40 |
|---|---|---|
| 30% slab, equity exemption off (the headline) | 9.32% | 11.41% |
| 30% slab, equity exemption on (Rs 10 lakh start) | 9.37% | 11.44% |
| 20% slab | 9.60% | 11.42% |
How sure we can be
One 21-year history is one draw. These numbers say how much the result could move with a different ordering of the same months, and how it looks once the 136 portfolios tried are taken into account.
- Return, 90% interval
- 7.3% to 14.3% a year before tax (block bootstrap of monthly returns; the point estimate is 10.7%).
- Against 60/40, 90% interval
- −4.6 pts to +2.3 pts a year. The interval straddles zero: the history cannot tell this portfolio and 60/40 apart.
- Five-year windows ahead of 60/40
- 27% of rolling five-year windows. The windows overlap: only about 4 of them are independent, so this is a description of the past, not a probability.
- Longest stretch behind 60/40
- 5,429 days
- Deflated Sharpe against 60/40
- 0.00. This is the probability that its edge over 60/40 is real once 136 tries are allowed for; 0.95 would be the usual bar.
- Worst five years after inflation
- −4.2% a year
- Rolling five-year return
- Lowest 5.2%, middle half 6.7% to 11.7%, highest 16.6% a year (before tax).
As a monthly SIP
Rs 10,000 a month, bought at the first session of each month, held for a fixed number of years and then sold, after costs and tax (with the annual equity exemption on). Each start month is one window. Windows that start a month apart share almost all their months, so the number of independent periods is printed next to every figure.
10-year SIP: after-tax return by start month
XIRR of each 10-year SIP, by the month it started (139 windows). Grey: the same instalments in one-year fixed deposits, taxed each year. FD proxy: 364-day T-bill primary yield (91-day before April 2005), after slab tax and cess each year.
How to read it. Each point is a whole 10-year SIP. Neighbouring points share most of their months, so the line is far smoother than the evidence is deep. The best start was Mar 2016 (11.5%), the worst Jul 2012 (5.5%).
per cent a year, after tax
The latest 10-year SIP in the data started in Oct 2016. It put in ₹12,00,000 and was worth ₹19,83,955 after tax at the end, against ₹14,81,288 for the same instalments in fixed deposits.
| Years | Windows | Independent periods | 5th pct | Median | 95th pct | Worst | Worst start | Below FD | Below inflation | Median value | More |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 3 | 223 | 7 | 2.2% | 7.2% | 15.8% | −5.0% | 2005-11 | 19% | 33% | ₹4,01,436 | |
| 5 | 199 | 4 | 4.2% | 7.2% | 13.7% | 3.3% | 2011-03 | 11% | 24% | ₹7,18,745 | |
| 7 | 175 | 3 | 5.2% | 6.6% | 11.8% | 4.6% | 2011-07 | 7% | 21% | ₹10,63,325 | |
| 10 | 139 | 2 | 5.8% | 6.9% | 10.3% | 5.5% | 2012-07 | 0% | 12% | ₹17,15,497 | |
| 15 | 79 | 1 | 6.4% | 7.1% | 9.0% | 6.2% | 2007-10 | 0% | 0% | ₹31,59,687 | |
| 20 | 19 | 1 | 7.9% | 8.5% | 9.1% | 7.8% | 2005-04 | 0% | 0% | ₹61,04,132 |
The workings
Data
Every chart has a table view, a CSV download and a link to its published JSON. Bundles are listed in data status.