What was tested, the main results and where every number comes from.
- Under the mandate the four covariance estimators realise 78.9 to 83.6 basis points a year of tracking error, and each one's forecast understates it. The bias statistic, realised over forecast, is 1.48 to 1.85, above the band of 0.8 to 1.25 fixed before the run.
- Index changes force 2.99 per cent a year of the Ledoit-Wolf path's 26.26 per cent of one-way turnover, or 11.4 per cent. On the Ledoit-Wolf estimator, loosening the cap on chosen turnover from 0.5 per cent a month to none raises it from 6.3 to 59.4 per cent a year, while the realised tracking error stays between 76.3 and 79.4 basis points a year. On the factor estimator chosen turnover rises from 6.2 to 14.9 per cent a year and the tracking error falls from 83.3 to 79.6.
- Month by month, allocation is the largest of the three Brinson-Fachler terms on the Ledoit-Wolf, factor and principal-component estimators: 16.4 to 16.7 basis points a month in absolute value on average, against 10.6 to 13.8 for selection and 1.5 to 2.2 for interaction. Over the 306 months the active return is 0.2 to 13.7 basis points a year before costs, within one Newey-West standard error of zero on every path, as is each of its three terms, and the six factors explain 2 to 13 per cent of its variance.
- The benchmark rebuilt from CRSP's list of index members reproduces CRSP's own value-weighted return of the same list with a tracking error of 11.5 basis points a year from 1926 to 2024. Its momentum and variance spreads have correlations of 0.888 and 0.954 with Kenneth R. French's large-cap spreads.1
Figure 1 | Forecast and realised tracking error under the mandate, four covariance estimators, July 1999 to December 2024
Each estimator's mean forecast (hollow) beside the tracking error it realised (filled), in basis points a year, with the bias statistic.
Every realised value lies to the right of its forecast. The four estimators realise 78.9 to 83.6 basis points a year, and the bias statistic, 1.48 to 1.85, lies above the band of 0.8 to 1.25 on all four. The forecast is the tracking error the month's own covariance matrix implies for the weights the optimiser chose.
| Estimator | Mean forecast, bp a year | Realised, bp a year | Bias statistic | Months |
|---|---|---|---|---|
| Ledoit-Wolf | 55.2 | 78.9 | 1.48 | 306 |
| Sample | 52.8 | 80.6 | 1.60 | 306 |
| Factor | 45.1 | 80.5 | 1.85 | 306 |
| Principal-component | 48.0 | 83.6 | 1.73 | 306 |
Simulated. Monthly rebalancing. Before fees. Bias statistic: the standard deviation of each month's active return divided by that month's forecast, over the months with a forecast, 1 for a calibrated forecast.
Source: CRSP and Compustat via Wharton Research Data Services, Kenneth R. French Data Library and author's calculations. Monthly, July 1999 to December 2024, US large caps, USD.
Table 1 | The four gates and nine expectations, fixed before each run, and what the run found
A gate not met stops the study. An expectation not met is reported as a finding, in the same words as one met. An expectation named for two or more estimators has one verdict, met only if it holds for each.
| Fixed before the run | What the run found | Verdict | |
|---|---|---|---|
| G1 | Tracking error of the benchmark against CRSP's own return of the same list, at most 50 bp a year | 11.5 bp a year | Met |
| G2 | Usable window at least 15 years | 99 years | Met |
| G3 | At least 90 per cent of removals with a CRSP price within 10 trading days | 100 per cent of 836 | Met |
| G4 | Capitalisation in at least 90 per cent of membership months for at least 90 per cent of names | 99.5 per cent of 1,912 names | Met |
| E1 | That tracking error at most 20 bp a year and the mean difference within 2 bp, with dividends and price-only | 11.5 and 11.6 bp a year, −0.18 and −0.16 bp | Met |
| E2 | Count within 10 names of CRSP's in every month, value within 2 per cent in at least 95 per cent of months | 1,188 and 1,186 of 1,188 months | Met |
| E3 | Momentum and variance spreads: correlation at least 0.80 with French's, mean difference within two standard errors | 0.888 and 0.954, 1.15 and 0.29 standard errors | Met |
| E4, first clause | Sample condition number at least 10 times the Ledoit-Wolf one in every month | 86 of 306 months | Not met |
| E4, second clause | Under the mandate, the sample estimator's realised tracking error above the Ledoit-Wolf one | 80.6 against 78.9 bp a year | Met |
| E5 | On the frontier, the mean forecast does not rise as the cap loosens, and the realised tracking error is lower at 4 per cent than at 0.5 | holds for the factor estimator, fails for Ledoit-Wolf (79.0 against 76.3 bp a year) | Not met |
| E6 | Realised tracking error 50 to 150 bp a year and bias statistic 0.8 to 1.25, each structured estimator | tracking error 78.9 to 83.6, bias 1.48 to 1.85 | Not met |
| E7 | Forced turnover less than a third of the total, Ledoit-Wolf path | 2.99 of 26.26 per cent a year, 11.4 per cent | Met |
| E8 | The three terms sum to the active return, and allocation is the largest in absolute value and interaction the smallest, each structured estimator | largest gap 9.6e-17, allocation 16.4 to 16.7, selection 10.6 to 13.8, interaction 1.5 to 2.2 bp a month | Met |
| E9 | Contributions and residual sum to the active return, and the mean exposure to the market within 0.05 of zero on every path | largest gap 8.7e-19, exposure −0.0006 to +0.0001 | Met |
Source: CRSP and Compustat via Wharton Research Data Services, Kenneth R. French Data Library and author's calculations. Monthly, January 1926 to December 2024, US large caps, USD.
Whether the rebuilt benchmark and the four covariance matrices are what they claim to be.
The universe in a month is the stocks on CRSP's list of index members on the last trading day of the month before, each weighted by its market capitalisation at the start of the month. A delisting is applied as Shumway (1997)4 proposes, and a member's sector is the GICS sector in force on the same day, through CRSP's link to Compustat. The index as S&P publishes it is not the benchmark: its total return is not in the extraction, and its float adjustment cannot be rebuilt from CRSP's share counts.
Table 2 | The benchmark's checks, January 1926 to December 2024
The benchmark against CRSP's own figures and the universe's sorts against French's, with the conditioning of the covariance matrices.
| Measure | Figure | Period |
|---|---|---|
| Benchmark against CRSP's value-weighted return, with dividends | 11.5 bp a year | January 1926 to December 2024 |
| The same, price-only | 11.6 bp a year | January 1926 to December 2024 |
| The same, with dividends, over the mandate's window | 15.2 bp a year | July 1999 to December 2024 |
| Benchmark's price return against the index's own, no verdict | 93.6 bp a year | to December 2004 |
| The same, from January 2005, when the index weights by float | 22.4 bp a year | January 2005 to December 2024 |
| Member-months in the universe | 440,724 | January 1926 to December 2024 |
| Momentum spread against French's: correlation | 0.888 | 1,172 months from January 1927 |
| Momentum: mean monthly difference (standard error) | −10.4 (9.0) bp | |
| Variance spread against French's: correlation | 0.954 | 417 months from April 1990 |
| Variance: mean monthly difference (standard error) | 2.7 (9.4) bp | |
| Months with the sample condition number at least 10 times the Ledoit-Wolf one | 86 of 306 | all from 2017 |
| The same with one share class for each company, no verdict | 30 of 306 | all from 2020 to 2022 |
The months in which the sample matrix's condition number reaches ten times the Ledoit-Wolf one have a cause found after the run: from 2014 the universe holds up to five companies with two share classes each, whose daily returns move almost together, and the difference between two such classes sets the sample matrix's smallest eigenvalue.
Source: CRSP and Compustat via Wharton Research Data Services, Kenneth R. French Data Library and author's calculations. Monthly, January 1926 to December 2024, US large caps, USD.
Table 3 | The condition numbers of the four covariance matrices by year, July 1999 to December 2024
The median over each year's months. A larger number means small errors in the matrix become larger errors in the weights an optimiser builds on it.
| Year | Sample | Ledoit-Wolf | Factor | Principal-component | Sample over Ledoit-Wolf |
|---|---|---|---|---|---|
| 1999 | 8,176 | 1,700 | 714 | 756 | 4.8 |
| 2000 | 8,609 | 1,626 | 658 | 723 | 5.3 |
| 2001 | 10,806 | 2,142 | 789 | 847 | 5.0 |
| 2002 | 14,485 | 2,826 | 1,151 | 1,174 | 5.1 |
| 2003 | 21,580 | 4,482 | 1,421 | 1,438 | 4.8 |
| 2004 | 26,928 | 5,054 | 1,722 | 1,777 | 5.3 |
| 2005 | 23,212 | 4,341 | 1,611 | 1,641 | 5.3 |
| 2006 | 11,496 | 3,066 | 1,008 | 1,088 | 3.7 |
| 2007 | 11,387 | 2,888 | 909 | 1,038 | 3.9 |
| 2008 | 16,719 | 3,737 | 1,241 | 1,374 | 4.5 |
| 2009 | 53,536 | 8,537 | 3,511 | 4,337 | 6.3 |
| 2010 | 55,505 | 10,177 | 3,708 | 4,655 | 5.5 |
| 2011 | 71,000 | 10,633 | 3,844 | 4,718 | 6.7 |
| 2012 | 47,285 | 13,388 | 2,955 | 3,397 | 3.5 |
| 2013 | 36,142 | 11,167 | 2,510 | 2,744 | 3.2 |
| 2014 | 31,841 | 9,292 | 2,203 | 2,362 | 3.4 |
| 2015 | 18,969 | 5,364 | 1,383 | 1,320 | 3.5 |
| 2016 | 36,119 | 6,573 | 1,682 | 1,723 | 5.5 |
| 2017 | 66,489 | 6,067 | 1,662 | 1,757 | 11.0 |
| 2018 | 69,934 | 4,947 | 1,573 | 1,672 | 14.1 |
| 2019 | 90,050 | 4,185 | 1,122 | 1,140 | 21.5 |
| 2020 | 380,906 | 4,996 | 2,724 | 2,445 | 76.2 |
| 2021 | 307,945 | 5,365 | 2,798 | 2,580 | 57.4 |
| 2022 | 171,875 | 5,850 | 3,340 | 3,299 | 29.4 |
| 2023 | 94,679 | 7,870 | 2,014 | 2,132 | 12.0 |
| 2024 | 101,795 | 9,341 | 1,725 | 1,961 | 10.9 |
Source: CRSP via Wharton Research Data Services, Kenneth R. French Data Library and author's calculations. Monthly, each from three years of daily returns, July 1999 to December 2024, US large caps, USD.
How much tracking error each estimator forecast and realised under the mandate, what the turnover cap buys and what the turnover is made of.
Each month the mandate chooses the weights that make the forecast tracking error as small as its constraints allow. Twenty-four paths run from the benchmark over the 306 months: the four estimators under four cumulative constraint sets at the 2 per cent cap, and the Ledoit-Wolf and factor estimators under the mandate at four further caps. Under the first three sets nothing pushes the portfolio off the benchmark, and it stays there in every month. The figures below are the twelve paths under the mandate.
Figure 2 | Realised and forecast tracking error by year, the Ledoit-Wolf estimator under the mandate, 1999 to 2024
The tracking error realised in each calendar year and the mean of that year's monthly forecasts, in basis points a year. The year 1999 holds six months.
The realised tracking error runs from 33 basis points a year in 2013 to 176 in 2008, when the mean forecast was 101. The realised tracking error is above the mean forecast in 22 of the 26 calendar years, and not in 2009, 2010, 2011 and 2013, so the understatement is not confined to one crisis.
| Year | Months | Realised, bp a year | Mean forecast, bp a year |
|---|---|---|---|
| 1999 | 6 | 86.3 | 46.3 |
| 2000 | 12 | 82.1 | 48.9 |
| 2001 | 12 | 87.6 | 72.4 |
| 2002 | 12 | 76.9 | 74.3 |
| 2003 | 12 | 85.0 | 67.8 |
| 2004 | 12 | 78.2 | 59.3 |
| 2005 | 12 | 100.2 | 66.1 |
| 2006 | 12 | 92.9 | 71.7 |
| 2007 | 12 | 92.5 | 75.5 |
| 2008 | 12 | 176.3 | 101.3 |
| 2009 | 12 | 90.4 | 103.3 |
| 2010 | 12 | 65.7 | 81.8 |
| 2011 | 12 | 72.7 | 81.1 |
| 2012 | 12 | 60.6 | 52.8 |
| 2013 | 12 | 33.3 | 44.6 |
| 2014 | 12 | 55.3 | 40.1 |
| 2015 | 12 | 62.4 | 33.0 |
| 2016 | 12 | 65.2 | 36.4 |
| 2017 | 12 | 46.5 | 34.8 |
| 2018 | 12 | 58.2 | 34.0 |
| 2019 | 12 | 48.3 | 33.4 |
| 2020 | 12 | 70.3 | 28.4 |
| 2021 | 12 | 38.0 | 27.6 |
| 2022 | 12 | 68.2 | 41.3 |
| 2023 | 12 | 47.2 | 42.6 |
| 2024 | 12 | 33.5 | 31.5 |
Simulated. Monthly rebalancing. Before fees. The realised figure of a year is the standard deviation of its monthly active returns, times the square root of 12.
Source: CRSP and Compustat via Wharton Research Data Services, Kenneth R. French Data Library and author's calculations. Monthly, July 1999 to December 2024, US large caps, USD.
Figure 3 | Tracking error against chosen turnover at five caps, the Ledoit-Wolf and factor estimators under the mandate, July 1999 to December 2024
For each estimator and each cap on chosen turnover (0.5, 1, 2 and 4 per cent a month, and none): the realised (filled) and mean forecast (hollow) tracking error in basis points a year, against the chosen turnover in per cent a year. The dots of each estimator run from the tightest cap on the left to no cap on the right.
E5 is not met. On both estimators the mean forecast falls as the cap loosens, from 64.5 to 54.1 basis points a year on Ledoit-Wolf and from 46.6 to 45.0 on the factor estimator. The realised tracking error at the 4 per cent cap is 79.6 on the factor estimator, below its 83.3 at 0.5 per cent, and 79.0 on Ledoit-Wolf, above its 76.3. On Ledoit-Wolf chosen turnover runs from 6.3 to 59.4 per cent a year across the caps while the realised tracking error stays between 76.3 and 79.4. The factor estimator chooses at most 14.9 per cent a year, and the 2 per cent cap binds on it in 13.7 per cent of months, against every month on Ledoit-Wolf. In a month where the other constraints need more chosen turnover than the cap allows, the cap gives way to the least turnover they need. This happens in the first month of every capped path under the mandate, July 1999, when the portfolio leaves the benchmark for the mandate and needs 6.41 per cent, and in 6 further months at the 0.5 per cent cap on the Ledoit-Wolf estimator and 3 on the factor estimator. So on the Ledoit-Wolf estimator at caps of 0.5, 1 and 2 per cent, where the cap binds in every month, chosen turnover a year lies a little above twelve times the cap.
| Estimator and cap | Chosen turnover, % a year | Total turnover, % a year | Realised, bp a year | Mean forecast, bp a year |
|---|---|---|---|---|
| Ledoit-Wolf, 0.5 | 6.3 | 8.8 | 76.3 | 64.5 |
| Ledoit-Wolf, 1 | 12.2 | 14.6 | 78.4 | 58.3 |
| Ledoit-Wolf, 2 | 24.2 | 26.3 | 78.9 | 55.2 |
| Ledoit-Wolf, 4 | 45.7 | 47.4 | 79.0 | 54.3 |
| Ledoit-Wolf, none | 59.4 | 61.0 | 79.4 | 54.1 |
| Factor, 0.5 | 6.2 | 8.6 | 83.3 | 46.6 |
| Factor, 1 | 10.5 | 12.6 | 82.3 | 45.3 |
| Factor, 2 | 13.2 | 15.1 | 80.5 | 45.1 |
| Factor, 4 | 14.6 | 16.5 | 79.6 | 45.0 |
| Factor, none | 14.9 | 16.7 | 79.6 | 45.0 |
Simulated. Monthly rebalancing. Before fees. Chosen turnover is one-way and excludes the turnover that index changes force.
Source: CRSP and Compustat via Wharton Research Data Services, Kenneth R. French Data Library and author's calculations. Monthly, July 1999 to December 2024, US large caps, USD.
Table 4 | The twelve paths under the mandate, July 1999 to December 2024
Tracking error, active return and its standard error in basis points a year. Turnover one-way, in per cent a year.
| Path | Realised tracking error | Mean forecast | Bias statistic | Active return, gross | Newey-West standard error | Active return, net at 10 bp | Total turnover, % a year | Chosen turnover, % a year | Names held, mean |
|---|---|---|---|---|---|---|---|---|---|
| Sample, cap 2 per cent | 80.6 | 52.8 | 1.60 | +13.7 | 18.6 | +11.1 | 26.3 | 24.2 | 375 |
| Ledoit-Wolf, cap 2 per cent | 78.9 | 55.2 | 1.48 | +12.7 | 18.2 | +10.1 | 26.3 | 24.2 | 388 |
| Factor, cap 2 per cent | 80.5 | 45.1 | 1.85 | +8.7 | 17.5 | +7.1 | 15.1 | 13.2 | 473 |
| Principal-component, cap 2 per cent | 83.6 | 48.0 | 1.73 | +0.2 | 17.4 | −1.7 | 19.3 | 17.3 | 459 |
| Ledoit-Wolf, cap 0.5 per cent | 76.3 | 64.5 | 1.26 | +7.2 | 17.2 | +6.3 | 8.8 | 6.3 | 427 |
| Ledoit-Wolf, cap 1 per cent | 78.4 | 58.3 | 1.39 | +13.4 | 18.2 | +11.9 | 14.6 | 12.2 | 404 |
| Ledoit-Wolf, cap 4 per cent | 79.0 | 54.3 | 1.52 | +10.1 | 18.3 | +5.4 | 47.4 | 45.7 | 383 |
| Ledoit-Wolf, cap none | 79.4 | 54.1 | 1.53 | +11.3 | 18.9 | +5.2 | 61.0 | 59.4 | 381 |
| Factor, cap 0.5 per cent | 83.3 | 46.6 | 1.82 | +7.2 | 17.5 | +6.3 | 8.6 | 6.2 | 474 |
| Factor, cap 1 per cent | 82.3 | 45.3 | 1.85 | +8.7 | 17.6 | +7.4 | 12.6 | 10.5 | 474 |
| Factor, cap 4 per cent | 79.6 | 45.0 | 1.85 | +8.4 | 17.5 | +6.7 | 16.5 | 14.6 | 473 |
| Factor, cap none | 79.6 | 45.0 | 1.85 | +8.5 | 17.5 | +6.8 | 16.7 | 14.9 | 473 |
Simulated. Monthly rebalancing. Gross of fees. The net active return is the gross less a trading cost of 10 basis points per unit of one-way turnover, forced turnover included. No active return is beyond one Newey-West standard error of zero.
Source: CRSP and Compustat via Wharton Research Data Services, Kenneth R. French Data Library and author's calculations. Monthly, July 1999 to December 2024, US large caps, USD.
Table 5 | What the turnover is made of, July 1999 to December 2024
One-way turnover in per cent a year, split into the part index changes force, the part that follows the benchmark's own reweighting and the part the optimiser chooses. The three parts less their overlap equal the total.
| Path | Total | Forced | Reweighting | Rebalancing | Overlap |
|---|---|---|---|---|---|
| The benchmark held (Ledoit-Wolf under C0) | 5.78 | 2.80 | 3.71 | 0.00 | 0.73 |
| Ledoit-Wolf, mandate at 2 per cent | 26.26 | 2.99 | 3.67 | 24.94 | 5.34 |
| Sample, mandate at 2 per cent | 26.29 | 3.00 | 3.67 | 25.13 | 5.51 |
| Factor, mandate at 2 per cent | 15.14 | 2.88 | 3.65 | 11.91 | 3.31 |
| Principal-component, mandate at 2 per cent | 19.29 | 2.93 | 3.69 | 16.48 | 3.82 |
E7 is met. On the Ledoit-Wolf path forced turnover is 2.99 per cent a year of a total of 26.26, or 11.4 per cent, below the third fixed before the run. A portfolio that holds the benchmark turns over 5.78 per cent a year: forced turnover 2.80, reweighting 3.71 and no rebalancing.
Source: CRSP and Compustat via Wharton Research Data Services, Kenneth R. French Data Library and author's calculations. Monthly, July 1999 to December 2024, US large caps, USD.
Where the mandate's active return came from, by sector and by factor.
Each month the active return of the four paths at the 2 per cent cap is split by the twelve sectors with the method of Brinson and Fachler (1985)2 into allocation, selection and interaction, which sum to it exactly, and by six factors into contributions and a residual. Both sides are taken over the month's names with a return. A sector with no stock above a weight of one millionth has a selection of zero and keeps the rest of its part as interaction. The monthly terms are summed over months with no geometric linking.
Figure 4 | Allocation and selection by sector over the window, the Ledoit-Wolf path under the mandate, July 1999 to December 2024
Each sector's allocation and selection as annual averages, in basis points a year, each with a bar of two Newey-West standard errors on each side.
On this path the active return of 12.7 basis points a year splits into allocation 1.4, selection 11.3 and interaction 0.0, with Newey-West standard errors of 15.9, 14.0 and 2.0. In Energy and Utilities, the two tilted sectors, selection is within a tenth of a basis point a year of zero: the tilt caps each stock at half its benchmark weight, the cap binds on 19,431 of the 19,433 stock-months in the two sectors on this path, and a sector whose stocks all sit at the same fraction of their benchmark weights earns the benchmark sector's return. Their allocation is 1.0 and 2.9 basis points a year, on standard errors of 13.4 and 4.2. On this path 3 of the 36 sector terms, interaction included, which the figure does not show, lie two standard errors or more from zero, and 1 to 4 on the other paths, against, at the 5 per cent level, 1.8 on average if the tests were independent.
| Sector | Allocation | Standard error | Selection | Standard error | Interaction | Standard error |
|---|---|---|---|---|---|---|
| Energy | +0.98 | 13.42 | 0.00 | 0.00 | 0.00 | 0.00 |
| Materials | +1.03 | 2.83 | −0.19 | 3.01 | +0.32 | 1.58 |
| Industrials | −0.09 | 0.63 | +1.87 | 6.00 | −0.64 | 0.28 |
| Consumer Discretionary | +0.32 | 1.07 | +0.16 | 3.99 | −0.02 | 0.19 |
| Consumer Staples | +0.26 | 1.98 | −5.13 | 4.56 | −0.87 | 0.37 |
| Health Care | −0.47 | 2.18 | +3.67 | 2.98 | +0.31 | 0.34 |
| Financials | −1.60 | 2.41 | +6.92 | 4.52 | +0.40 | 0.31 |
| Information Technology | +1.85 | 1.17 | +3.40 | 4.44 | −0.05 | 0.18 |
| Communication Services | −1.16 | 1.24 | +0.10 | 1.39 | +0.04 | 0.17 |
| Utilities | +2.87 | 4.23 | 0.00 | 0.00 | 0.00 | 0.00 |
| Real Estate | −2.13 | 1.03 | −0.40 | 0.68 | −0.12 | 0.16 |
| Unclassified | −0.45 | 1.36 | +0.92 | 0.85 | +0.61 | 0.63 |
Simulated. Monthly, summed arithmetically. Gross of fees. Interaction is in the table.
Source: CRSP and Compustat via Wharton Research Data Services and author's calculations. Monthly, July 1999 to December 2024, US large caps, USD.
Table 6 | The three terms over the window, four paths under the mandate at the 2 per cent cap, July 1999 to December 2024
Annual averages with Newey-West standard errors, and E8's mean absolute monthly terms, each month's figure the sum over the twelve sectors.
| Path | Allocation, bp a year | Standard error | Selection, bp a year | Standard error | Interaction, bp a year | Standard error | Allocation, mean absolute, bp a month | Selection, mean absolute, bp a month | Interaction, mean absolute, bp a month | Sector-months not held |
|---|---|---|---|---|---|---|---|---|---|---|
| Ledoit-Wolf | +1.4 | 15.9 | +11.3 | 14.0 | 0.0 | 2.0 | 16.56 | 13.81 | 1.76 | 16 |
| Sample, no verdict | +1.3 | 16.1 | +12.3 | 14.7 | +0.1 | 2.2 | 16.63 | 14.37 | 1.86 | 22 |
| Factor | +5.5 | 16.3 | +4.0 | 9.3 | −0.8 | 1.5 | 16.74 | 10.63 | 1.50 | 0 |
| Principal-component | +3.5 | 15.4 | −3.9 | 9.6 | +0.6 | 2.5 | 16.41 | 12.33 | 2.21 | 0 |
E8 is met. The three terms sum to the active return in every month of the four paths, and on the three structured estimators allocation is the largest in absolute value and interaction the smallest. A sector is not held in 22 months on the sample estimator's path and in 16 on the Ledoit-Wolf path, all of them from 2022, and 66 member-months have no return and are left out of both sides.
Source: CRSP and Compustat via Wharton Research Data Services and author's calculations. Monthly, July 1999 to December 2024, US large caps, USD.
Table 7 | The factor attribution, four paths under the mandate at the 2 per cent cap, July 1999 to December 2024
Contributions, their sum, the residual and the active return in basis points a year, each with its plain standard error in brackets, the factor share in per cent and the mean active exposures.
| Ledoit-Wolf | Sample | Factor | Principal-component | |
|---|---|---|---|---|
| Market, contribution | −1.7 (0.5) | −1.6 (0.5) | −1.4 (0.5) | −1.2 (0.7) |
| Size, contribution | +0.7 (1.2) | +0.6 (1.2) | +0.9 (1.2) | +2.1 (1.9) |
| Value, contribution | −2.0 (2.5) | −1.9 (2.5) | −1.8 (1.5) | −5.4 (4.6) |
| Profitability, contribution | −4.7 (3.5) | −4.9 (3.4) | −2.3 (2.4) | −6.5 (5.5) |
| Investment, contribution | −1.3 (1.5) | −1.1 (1.5) | −0.8 (1.3) | −1.7 (2.5) |
| Momentum, contribution | −0.8 (2.2) | −0.7 (2.1) | 0.0 (1.7) | +0.7 (1.8) |
| The six together | −9.7 (4.4) | −9.6 (4.4) | −5.4 (3.3) | −12.0 (5.7) |
| Residual | +22.4 (14.9) | +23.4 (15.3) | +14.1 (15.8) | +12.3 (15.4) |
| Active return | +12.7 (15.6) | +13.7 (16.0) | +8.7 (15.9) | +0.2 (16.5) |
| Factor share, per cent | 8.9 | 8.6 | 2.1 | 12.9 |
| Market, mean active exposure | −0.0006 | −0.0006 | +0.0001 | −0.0002 |
| Size, mean active exposure | +0.0055 | +0.0053 | +0.0054 | +0.0056 |
| Value, mean active exposure | −0.0068 | −0.0067 | −0.0019 | −0.0100 |
| Profitability, mean active exposure | −0.0071 | −0.0071 | −0.0041 | −0.0162 |
| Investment, mean active exposure | +0.0045 | +0.0047 | +0.0006 | +0.0002 |
| Momentum, mean active exposure | −0.0028 | −0.0028 | −0.0023 | −0.0018 |
E9 is met. Contributions and residual sum to the active return in every month, and the mean active exposure to the market factor lies between −0.0006 and +0.0001 on the four paths. On the Ledoit-Wolf path the six contributions sum to −9.7 basis points a year and the residual to 22.4. The standard errors of this table are the plain ones the copied attribution computes, with no allowance for correlation from one month to the next.
Source: CRSP and Compustat via Wharton Research Data Services, Kenneth R. French Data Library and author's calculations. Monthly, July 1999 to December 2024, US large caps, USD.
The data, what is known at each date, the costs, what the study cannot show, and the terms.
Method
Data and access. CRSP and Compustat via Wharton Research Data Services, accessed through Queen Mary University of London, from an extraction of 28 September 2026. They are licensed, and the repository holds none of them: it holds code, tests, the extraction's SHA-256 and aggregate tables.5 Nothing that reproduces CRSP's index file is committed by month, and no weight of a security or of a sector of the portfolio is committed in any form. Kenneth R. French's files are downloaded at run time.
What is known at each date. Membership, sector and capitalisation are read on the last trading day of the month before. Each covariance matrix uses the three years of daily returns before the month. Nothing from the month itself enters a weight.
Rebalancing and costs. Monthly. The turnover cap applies to chosen turnover, measured from the portfolio left after last month's returns with leavers sold and entrants bought at their benchmark weight. Returns are gross of fees. A trading cost of 10 basis points per unit of one-way turnover is charged where stated, forced turnover included.
What the study cannot show. It is a simulation. The benchmark weights CRSP's list by full capitalisation, and the index weights by float. A GICS sub-industry is wider than a list of controversial weapons. Costs are proportional to turnover, with no market impact. Sectors exist from July 1999, so the mandate starts there. No active return is distinguishable from zero, so the study makes no claim about a return the mandate earns.
Where each exhibit comes from
The committed tables each exhibit is drawn from, each linked on GitHub at the commit that last changed it, and the notebooks that write them.
| Exhibit | Committed tables | Written by |
|---|---|---|
| Figure 1 | Forecast and realised tracking error under the mandate, four covariance estimators, July 1999 to December 2024 | paths_summary.json | notebook 09 |
| Table 1 | The four gates and nine expectations, fixed before each run, and what the run found | verdicts_milestone1.json, verdicts_milestone2.json, verdicts_milestone3.json, verdicts_milestone4.json, src/slt/constants.py | notebook 04, notebook 06, notebook 07, notebook 09, notebook 10, notebook 11 |
| Table 2 | The benchmark's checks, January 1926 to December 2024 | reconstruction_summary.json, universe_summary.json, answer_key_summary.json, covariance_summary.json | notebook 02, notebook 04, notebook 06, notebook 07 |
| Table 3 | The condition numbers of the four covariance matrices by year, July 1999 to December 2024 | covariance_by_year.csv | notebook 07 |
| Figure 2 | Realised and forecast tracking error by year, the Ledoit-Wolf estimator under the mandate, 1999 to 2024 | paths_by_year.csv | notebook 09 |
| Figure 3 | Tracking error against chosen turnover at five caps, the Ledoit-Wolf and factor estimators under the mandate, July 1999 to December 2024 | paths_summary.json, cap_relaxed.csv | notebook 09, notebook 10 |
| Table 4 | The twelve paths under the mandate, July 1999 to December 2024 | paths_summary.json | notebook 09 |
| Table 5 | What the turnover is made of, July 1999 to December 2024 | paths_summary.json, verdicts_milestone3.json | notebook 09, notebook 10 |
| Figure 4 | Allocation and selection by sector over the window, the Ledoit-Wolf path under the mandate, July 1999 to December 2024 | brinson_summary.json, tilt_cap.csv | notebook 11 |
| Table 6 | The three terms over the window, four paths under the mandate at the 2 per cent cap, July 1999 to December 2024 | brinson_summary.json, attribution_counts_by_year.csv, verdicts_milestone4.json | notebook 11 |
| Table 7 | The factor attribution, four paths under the mandate at the 2 per cent cap, July 1999 to December 2024 | factor_attribution_summary.json, verdicts_milestone4.json | notebook 11 |
Terms
- universe
- the point-in-time S&P 500: the stocks on CRSP's list of index members on the last trading day of the month before.
- market capitalisation
- price times shares outstanding, at the end of the month before.
- benchmark
- the cap-weighted universe, each stock's weight its market capitalisation over the universe's total.
- active weight
- a weight in the portfolio minus the same weight in the benchmark.
- active return
- the portfolio's return minus the benchmark's, each over the month's names with a return.
- cap-weighted
- weighted by market capitalisation. CRSP's value-weighted return of the list is the same, computed by CRSP.
- value-weighted
- cap-weighted.
- basis point
- a hundredth of a percentage point.
- tracking error
- the standard deviation of the monthly active return, times the square root of 12. The forecast is the one the month's covariance matrix implies for the month's active weights, the realised the measured one.
- covariance matrix
- the table of the stocks' variances and of their covariances with each other.
- covariance estimator
- a way of estimating the covariance matrix from three years of daily returns: the sample covariance, the Ledoit-Wolf estimator6, which shrinks it towards a scaled identity matrix, a factor model on French's five factors and momentum, and a principal-component model with the benchmark as its first factor. The last three are the structured estimators.
- eigenvalue
- the variance of one of the uncorrelated portfolios a covariance matrix splits into.
- condition number
- a covariance matrix's largest eigenvalue over its smallest.
- band
- the range of the bias statistic within which a forecast counts as right, 0.8 to 1.25, fixed before the run.
- bias statistic
- the standard deviation of each month's active return divided by that month's forecast, over the months with a forecast above one millionth. It is 1 when the forecasts are right on average.
- mandate
- the full set of constraints the portfolio obeys: long-only, each stock and each company within 50 basis points of its benchmark weight, each untilted sector within 2 percentage points, chosen turnover capped, the tilt, two sub-industries excluded and a beta of one to the benchmark.
- beta
- the slope of a return on the benchmark's return.
- tilt
- every Energy and Utilities stock capped at half its benchmark weight.
- constraint set
- one of the cumulative sets C0 to C3: long-only, plus the bounds per stock, company and sector, plus the turnover cap, plus the rest of the mandate.
- path
- one covariance estimator under one constraint set at one turnover cap, run month by month from the benchmark.
- turnover
- a month's one-way trading as a share of the portfolio. Forced turnover is what index changes require, chosen turnover the rest.
- turnover cap
- the limit on chosen turnover, 2 per cent of the portfolio a month in the mandate, with 0.5, 1 and 4 per cent and no cap on the frontier.
- reweighting
- the turnover that following the benchmark's own changes in share counts and its dividends takes.
- rebalancing
- the optimiser's own moves. The overlap is the trades of one step that the next undoes.
- Newey-West standard error
- the standard error of a mean that allows each month to be correlated with the next twelve. A plain standard error makes no such allowance.
- attribution
- a split of the active return into parts that sum to it.
- allocation
- the part of the active return from holding a sector at another weight than the benchmark's: the sector's active weight times the benchmark sector's return less the benchmark's.
- selection
- the part from holding other stocks within a sector: the sector's benchmark weight times the portfolio sector's return less the benchmark sector's.
- interaction
- the sector's active weight times that same difference.
- Unclassified
- the twelfth sector, the members with no GICS sector, one to three securities a month.
- factor
- a common source of return: the market, size, value, profitability and investment factors of Fama and French (2015), and momentum.
- loading
- a stock's sensitivity to a factor, from a regression of its daily returns on the six factors over the three years before the month.
- active exposure
- the members' loadings on a factor weighted by the active weights. A factor's contribution is its active exposure times its return, the residual the active return less the contributions, and the factor share one minus the residual's variance over the active return's.
- Sharpe ratio
- in the card of Optimal versus naive diversification below, a portfolio's mean return above the one-month Treasury bill rate, over its standard deviation.
- look-ahead
- the use, in a decision dated at one time, of data published only later.
- gate
- a test of the data fixed in advance, which stops the study if not met.
- expectation
- a test of a result fixed in advance, reported as a finding whether met or not.
Notes
- Kenneth R. French Data Library: the files of the five factors and momentum, and of the portfolios formed on size and momentum and on size and variance.
- Brinson, G. P. and Fachler, N. (1985). Measuring non-U.S. equity portfolio performance. Journal of Portfolio Management 11(3), 73 to 76. https://doi.org/10.3905/jpm.1985.409005
- Fama, E. F. and French, K. R. (2015). A five-factor asset pricing model. Journal of Financial Economics 116(1), 1 to 22. https://doi.org/10.1016/j.jfineco.2014.10.010
- Shumway, T. (1997). The delisting bias in CRSP data. Journal of Finance 52(1), 327 to 340. https://doi.org/10.1111/j.1540-6261.1997.tb03818.x
- The repository: the notebooks, the tables committed in results/20260928/ and the constants fixed before each run.
- Ledoit, O. and Wolf, M. (2004). A well-conditioned estimator for large-dimensional covariance matrices. Journal of Multivariate Analysis 88(2), 365 to 411. https://doi.org/10.1016/S0047-259X(03)00096-4
Other projects
Optimal versus naive diversification. Under an index-tracking mandate on French's 49 industries, five covariance estimators understate their own tracking error, by 22 to 61 per cent. The replication of DeMiguel, Garlappi and Uppal (2009) reproduces 29 of the 30 Sharpe ratios that carry a band. Page, dashboard, note, code
Climate risk as an equity signal. An ownership graph linking 328 listed companies to the 5,115 tracked assets they own. Of 19 apparent ownership changes checked by hand, 12 had no transaction behind them. Page, note, code
Volatility-managed portfolios: a replication, the real-time critique, and the months after the sample. Volatility estimation and position sizing on Ken French's equity factors, backtested out of sample and without look-ahead, replicating Moreira and Muir (2017) and Cederburg, O'Doherty, Wang and Yan (2020). Scaled by the inverse of last month's realised variance, a factor has a Sharpe ratio above the original's beyond its 95 per cent interval for one of six factors, momentum, and the 136 months after Moreira and Muir's sample cannot tell any difference from zero. Page, dashboard, note, code
Rules-based portfolio. Runs a real-money portfolio under rules published before the first order: two UCITS ETFs at 70/30, a monthly cycle with pre-trade checks, look-through exposure and ex-ante risk on a public dashboard. Page, dashboard, guide, code