Indian high-liability stocks beat low-liability peers; factor models fail two sorts

High-liability Indian stocks beat low-liability peers by enough to warrant a closer look: the low-minus-high total-liabilities spread averages -0.5% a month, with a t of -2.25. That long-short result is in-sample and before costs. The authors' claim that their capital-structure model fits best is much harder to defend from their tables.

What did Somu and Barik measure?

Somu and Barik study Indian stocks from 2010 to 2024. Their premise is familiar: shareholders bear more financial risk when a firm has more liabilities relative to assets, and should therefore expect a higher return. The choice of liabilities measure gives the paper its interest. Indian studies often use debt/equity, which captures formal borrowing. Here, LD/TA measures long-term debt over total assets, while TL/TA measures total liabilities over total assets. TL/TA also includes operating liabilities such as trade payables.

Bloomberg supplies monthly data for 394 non-financial NSE 500 firms from January 2010 to December 2024. The authors rebalance each June using March fiscal-year accounts. Their quintile portfolios are equal-weighted; they do not state how the factor portfolios are weighted. Factors use 2x3 size-by-characteristic sorts. SMB captures size. LMH is low minus high P/B and serves as the value factor because book-to-market data were unavailable. LMXLDTA and LMXTLTA are each low minus high on their respective capital-structure ratios.

The authors fit three models to three sets of 25 size-by-characteristic double-sorted test portfolios. FF3fm uses the standard three factors. MFF3fm replaces LMH with both capital-structure factors, leaving four factors despite its name. EFF3fm includes all five. They argue that MFF3fm describes returns better than its rivals, citing higher R2 and lower GRS statistics. The Gibbons-Ross-Shanken test asks whether all 25 intercepts are jointly zero.

The spread comes from total liabilities

Only the broader balance-sheet measure produces a significant factor spread. The single sorts give little sign of a clean gradient: TL/TA quintiles return 0.012, 0.015, 0.016, 0.015, 0.015. LD/TA quintiles return 0.012, 0.017, 0.018, 0.013, 0.013. That hump leaves 0.1 percentage point a month between the bottom and top LD/TA quintiles.

The factor returns sharpen the distinction. LMXLDTA averages 0.002 with t 1.012. LMXTLTA averages -0.005 with t -2.249: high-TL/TA firms out-earned low-TL/TA firms. The paper's text gives both t-statistics plainly. Its abstract's "significant capital structure premium" rests on TL/TA alone. The practical implications say so too: "the significant negative capital structure premium indicates that the firms with high total liabilities to total assets earned higher returns".

Liabilities beyond long-term debt appear to carry the priced information. That category includes short-term borrowing and payables. The TL/TA factor is significant; the value factor is not. For their value claim, the authors cite single-sort quintiles, where the lowest-P/B quintile earns 0.020 against 0.010 for the highest. Yet LMH earns 0.003 with t 0.845, and its correlation with LMXTLTA is -0.671. Their text says "only the size premium is statistically significant" among market, size and value. The abstract, meanwhile, claims a "strong size, value, and capital structure effect".

The fit tables resist the ranking

MFF3fm does better on the count of significant intercepts. At 5%, FF3fm has 3, 2 and 4 across the size-value, size-LD/TA and size-TL/TA sorts. MFF3fm has 1, 1 and 1; EFF3fm has 2, 2 and 3. MFF3fm deserves credit for that result.

The joint tests give a less tidy ranking. Against FF3fm, MFF3fm has a lower GRS statistic on size-value (1.809 vs 1.921) and size-TL/TA (2.687 vs 2.850), though a higher one on size-LD/TA (1.420 vs 1.391). EFF3fm beats MFF3fm on size-value GRS (1.758 vs 1.809). EFF3fm also has higher average R2 on size-value (0.778 vs 0.766) and size-TL/TA (0.779 vs 0.772). The paper's GRS paragraph gets the F comparisons right, then says higher R2 corroborates them. On size-TL/TA, the comparison with EFF3fm runs the other way (0.772 vs 0.779).

The conclusion awards MFF3fm the win "for all three portfolio combinations" because "the resulting AAA values are lower and indistinguishable from zero, and the R2 values are higher". Its GRS claim is narrower: the statistics "are lower for MFF3fm in most of the combination". The broader verdict fails on its own measures. Average absolute alpha is 0.003 for all three models on size-value. EFF3fm has higher average R2 than MFF3fm on size-value (0.778 vs 0.766) and size-TL/TA (0.779 vs 0.772). The EFF3fm section says average R2 "declines in the case of all the combinations relative to MFF3fm". Its table says otherwise.

One table entry raises a more basic problem. EFF3fm adds LMH to every MFF3fm regressor. If OLS uses the same months, that addition cannot lower R2. Yet on size-LD/TA, EFF3fm reports 0.754 against MFF3fm's 0.767. The small, high-LD/TA portfolio falls from 0.715 to 0.654. With one extra regressor over roughly fifteen years of monthly observations, adjusted R2 could decline by only a fraction of a point. A fall from 0.715 to 0.654 is impossible under either definition. The samples may differ, or something may be misreported.

The EFF3fm size-value SMB block gives another reason to question the table. Its small-size row (1.453, 0.764, 0.499, 0.069, -0.159) nearly reproduces the FF3fm low-value column (1.455, 0.765, 0.496, 0.060, -0.162); it looks transposed. The abstract cites adjusted R2, whereas every table reports R2. The distinction matters when MFF3fm has four factors and FF3fm has three.

GRS still rejects

All three models fail the GRS test at 5% on two of the three portfolio sorts. Size-value p-values are 0.009, 0.020 and 0.022. Size-TL/TA p-values are 0.000 for every model. Size-LD/TA alone passes throughout at 5%, with 0.117, 0.104 and 0.064. It is also the sort on which capital structure barely prices.

The authors write that "a lower GRS F-statistics along with the higher average R2 marks the model more efficient". Here, a lower GRS statistic can amount to a lower degree of rejection. MFF3fm still rejects on the TL/TA sort, where its new factor ought to matter most, with F 2.687.

Some distance from a tradable book

The authors acknowledge a major limitation: "the study has not dealt with survivorship bias during the study period". They provide no sub-period analysis. Their NSE 500 constituent universe has no treatment of delistings, a concern that bears especially hard on equal-weighted small portfolios. The smallest size quintile earns 0.033 a month (t 4.906), versus 0.004 for the largest; SMB earns 0.014 (t 5.722). Those figures call for scrutiny in an equal-weighted index universe. We did not find turnover, transaction cost or liquidity figures anywhere in the paper. From 394 firms, the 25-portfolio sorts average under 16 stocks each.

We could not rebuild the results. Our data contains no Indian equities or Indian company fundamentals, so we cannot reconstruct the NSE 500 universe, the BSE-200 market proxy or the 91-day T-bill series used by the factors.

The TL/TA spread merits further work. It implicates liabilities beyond long-term debt, including short-term borrowing and payables. The model ranking does not merit repetition. I would change my view if the TL/TA spread held in 2017 to 2024 alone, value-weighted and with delisted names restored.