Guide · Valuation and rates · Updated 2026-09-23

An earnings yield is a real rate and a bond yield is a nominal one, so the gap between them measures inflation before it measures value

The rule fits on one chart: when the market’s earnings yield falls far below the government bond yield, equities are dear and later returns should disappoint. In India it is quoted with the Nifty 50’s P/E and the 10-year benchmark yield, and almost always without three facts: what the P/E is made of, what kind of rate each side is, and how little thirteen years of monthly data can say about five-year returns. Every figure below comes from the exchange’s daily index files, 3,385 sessions from 2013-01-01 to 2026-09-18, the OECD’s monthly 10-year yield for India and the official consumer price index, and the closing note gives the method in enough detail to redo it.

The bond yield reaches the price through the discount rate, and only part of it arrives

An index is a claim on the cash its companies will pay out, and its price is those payments discounted at the return investors require, k. Split k into the government bond yield and whatever extra investors demand for bearing equity risk, the equity risk premium, and the bond yield sits inside every equity valuation by construction. That is the discount-rate channel. If payments grow at a steady rate g, the sum collapses to price equals next year’s dividend divided by k minus g, and the sensitivity follows by differentiation: a change in k moves the log price by minus one over k minus g, which is the price divided by next year’s dividend. With the index’s dividend yield at 1.22 per cent in July 2026, that makes equity a very long bond on paper.

Hold growth fixed and raise the discount rate one point, and the dividend form cuts the price by 43 per cent; the earnings form, which treats the index as flat real earnings discounted at a real rate of 4.83 per cent, cuts it by 17 per cent. Measured against the 10-year yield, month-average index level on month-average yield across 162 monthly changes, the index moved -3.37 per cent per point, standard error 1.21, and the yield change explained 2.3 per cent of the index’s monthly movement.

The distance between the model rows and the measured one is the first lesson of the channel. A nominal yield moves for three reasons: expected inflation, the real rate and the term premium. Expected inflation raises k and g together, because revenues and costs are in the same rupees that inflation erodes, so k minus g and the price stay roughly where they were. Only the real part and any change in the equity premium itself press on the price, and the nominal yield does not say which part moved. A rule that treats every point of nominal yield as a point of discount rate over-reads the channel several times over.

The P/E is a ratio of sums, and its definition changed on 31 March 2021

The exchange computes the index P/E as index market capitalisation divided by gross earnings: each constituent’s earnings over the trailing four quarters, profits and losses both, adjusted for free float and capping and then summed (NSE, Price Earnings Ratio methodology page, updated 20 October 2025). Three things follow. It is a ratio of sums, so the largest weights dominate it and a large loss shrinks the denominator. It is trailing, so it divides today’s price by last year’s earnings. And its earnings are whatever the constituent list and the definition say they are on the day.

On 23 February 2021 the index maintenance sub-committee of NSE Indices changed that definition with effect from 31 March 2021: earnings moved from standalone to consolidated financials, falling back to standalone only where consolidated figures are unavailable, and the dividend yield moved from dividends reported in the annual report to equity dividends over a rolling twelve months by ex-date (NSE Indices press release, 23 February 2021). The session files show what that did.

The P/E fell 17.88 per cent while the index fell 1.04 per cent, so 94.7 per cent of the fall was the definition. On that one session’s arithmetic, consolidated earnings ran 20.5 per cent above standalone, a figure that also absorbs the one constituent replaced the same day. The dividend yield fell 10.3 per cent for the same reason, and the P/B, whose definition did not change, moved 0.2 per cent. Every chart of the index P/E that crosses March 2021 unadjusted, and every claim that today’s P/E sits above or below its long-run average, compares two different ratios. A long-run average quoted without its earnings basis cannot be checked.

Divide the index by its P/E and the result is the earnings the ratio divides by. Leaving out the switch, that figure moved by more than one per cent in a single session 111 times, and 81 per cent of those fell in the eight months when quarterly results arrive, which hold 67 per cent of sessions; listed companies file within 45 days of a quarter and 60 days of the year under Regulation 33 of the SEBI listing regulations. Some steps are composition: on 13 July 2023, when a constituent that had merged into another was replaced (NSE Indices press release, 4 July 2023), the P/E rose 5.2 per cent more than the index. And the trailing figure carries the earnings cycle. From December 2019 to 8 February 2021, when the P/E peaked at 42.0, the index rose 24.2 per cent and the earnings it was divided by fell 16.3 per cent. A trailing earnings yield at an earnings trough describes the trough as much as the price. The guide to the P/E ratio covers trailing against forward readings for a single company.

Two honest treatments of the break exist. Use the published series and state the break, or splice it: multiply every P/E before 31 March 2021 by the switch-day ratio of 0.8298. The splice assumes the gap between consolidated and standalone earnings stayed constant for eight years, which the files cannot confirm. Everything below is reported both ways, with the spliced series called one basis.

The gap, month by month, was below zero in all 163 months

Each month’s earnings yield here is the average of 100 divided by the P/E over every session in the month. The bond side is the OECD’s monthly long-term government bond yield for India (Main Economic Indicators, FRED series INDIRLTLT01STM), which article 155 found behaves as a monthly average, so averages are set against averages.

On the published P/E the gap averaged -2.83 points, from -4.38 in September 2018 to -1.20 in May 2020. On one basis it averaged -2.29, from -3.63 to -0.24 in the same two months. Read literally, the rule rated Indian equities expensive against government bonds in every month from January 2013 to July 2026. Over the calendar years 2015 to 2025 the index’s total return averaged 11.70 per cent a year in log terms, against 6.36 per cent for the exchange’s 10-year benchmark government bond index.

The bond series has its own seam. It follows whichever bond is the current 10-year benchmark and steps when that bond changes, while the exchange’s 10-year benchmark clean price index is chain-linked across the same changes. So the yield change the price index implies, at a modified duration of about seven, isolates the step. The two agree to within a median 2.5 basis points a month; 16 of 136 months differ by more than 10, the largest -62 basis points in May 2020. Undoing that step would take May 2020’s gap to -0.86, further below zero. The steps cannot be removed as level shifts, because summed they reach 2.24 points by 2026, which says a chain-linked price and a benchmark yield are different objects: a fixed bond against whichever bond is current. The predictive test below is therefore re-run with those 16 months dropped.

An earnings yield is a real return, and a bond yield is not

Take a company whose real earnings are flat and which pays them all out. Inflation lifts its rupee earnings at the inflation rate without any reinvestment, because its prices rise with everyone else’s. Its price is next year’s earnings divided by k minus inflation, so its earnings yield is k minus inflation: the real return investors require. The same holds for a company that reinvests, provided the reinvestment earns only the return investors require; where it earns more, the earnings yield understates the real return. A government bond has no such escalator; its yield is a real rate plus expected inflation. Subtract, and the yield gap is approximately the real equity premium minus expected inflation. The argument that investors wrongly capitalise real earnings at nominal rates is Modigliani and Cohn’s (Financial Analysts Journal, 1979); its application to this comparison, sometimes called the Fed model, is Asness’s (Journal of Portfolio Management, 2003).

Where inflation runs near 2 per cent, the bias is 2 points and a gap near zero can pass for fair value. India’s inflation target is 4 per cent on the consumer price index with a band of 2 to 6, set under section 45ZA of the RBI Act, 1934, and retained for the five years from 1 April 2026 to 31 March 2031 (RBI, monetary policy framework overview). On the target alone, a gap of minus 4 is what a zero premium looks like, and the measured average of -2.29 is consistent with a positive premium. The rule’s threshold is an inflation forecast in disguise.

Subtracting a constant such as the target moves the threshold and leaves every month’s rank unchanged, so it cannot help or hurt a predictive test. Inflation that varies does. The inflation-indexed bonds the government issued in 2013 were linked to wholesale prices, not consumer prices (RBI, FAQs on Inflation Indexed Bonds, 28 May 2013), so they could not supply a consumer-price real yield even while they traded. The real yield here is therefore built as the 10-year yield less the latest year-on-year change in the official consumer price index, All India, Combined (MoSPI, base 2012=100).

Corrected that way, the gap averaged +2.99 points and sat below zero in 13 of 156 months. In November 2013, with consumer prices 11.51 per cent above a year earlier on MoSPI’s index, the gap read -2.09 on one basis and +9.41 against a real yield, the highest in the record, and the index returned 34.5 per cent in log terms over the next twelve months. The correction has its own failure. Trailing inflation is not expected inflation: in October 2025 measured inflation fell to 0.25 per cent and the corrected gap dropped to -1.78, its lowest, for a reason no investor would call valuation.

The comparison also rests on a premise: that the earnings yield moves with the bond yield. In levels the two correlate at 0.73 across 163 months, which looks like confirmation. In twelve-month changes the correlation is 0.26. Two series that both drifted over the decade correlate in levels whether or not one responds to the other, which is the trap article 97 on stationarity describes; the change correlation is the one that tests the premise, and it is weak.

The predictive test, with the overlap counted

The claim is that a low gap predicts poor returns. For every month end from January 2013, take the index’s total return over the next one, three and five years, rebuilt from the price index and the exchange’s dividend points as in article 141, annualise it in log terms, and regress it on that month’s gap. That gives 152, 128 and 104 observations. It does not give that many pieces of evidence. Two five-year windows starting a month apart share 59 of their 60 months, and the record holds only 2.7 non-overlapping five-year windows, 4.6 three-year ones and 13.7 one-year ones.

Ordinary standard errors treat every observation as new, so they are too small by roughly the square root of the overlap, and the gap’s own persistence makes it worse: its month-to-month autocorrelation on one basis is 0.90. Newey-West errors repair the overlap in large samples and under-repair it in small ones. To measure by how much, the same regressions were run 10,000 times on simulated data in which the gap predicts nothing, keeping its persistence, the index’s own monthly mean and volatility, and the measured correlation of -0.47 between shocks to the gap and to returns, with numpy’s default generator and seed 160.

At one year the one-basis gap has a slope of +10.2, so a gap one point higher, about 1.7 standard deviations, went with a next-year return about 10 points higher; R² is 0.21 and the simulated p-value 0.046 (0.041 to 0.046 across ten seeds), borderline. At three years there is nothing, R² 0.01 and p 0.77. At five years the slope is +2.07 with a Newey-West t of 5.49, which reads as overwhelming, and the simulated p-value is 0.26: with fewer than three independent windows, a fit that tight arises by chance about one time in four. The null is not even centred on zero. A falling market raises the gap in the same month, and the gap is persistent, so under no predictability the fitted one-year slope still averages +1.94, in the direction the rule predicts; Stambaugh (Journal of Financial Economics, 1999) derived this bias.

The pieces point the same way as the real-and-nominal argument, without proving it. The bond yield alone has no one-year relationship with returns (R² 0.004), and the earnings yield alone does worse than the gap at one year and turns negative at three. The inflation-corrected gap has the tightest one-year fit, R² 0.30 with a simulated p of 0.015, and it too fades to nothing at three and five years (p 0.68 and 0.69). Fifteen tests were run, and two came in below 0.05. Were the tests independent, the chance of at least one doing so with nothing to find would be 54 per cent; they are not independent, which lowers that figure, but two survivors out of fifteen are not a discovery.

The one-year result also depends on which months are in it. Dropping the 2020 start months, when a crash and a rebound sat inside single windows, cuts R² from 0.21 to 0.13. The only stretch on a single published definition, April 2021 onward, has 53 windows and a slope of -11.3, the opposite sign. Using just one window a year gives 12 or 13 truly separate observations and twelve ways to pick the starting month; the slope then runs from +1.9 to +14.6, with t-statistics from 0.24 to 3.12, depending on nothing but that choice. The record supports a weak one-year association that leans on two episodes and a splice, and no measurable three- or five-year one. That is what testing against the right null means here: the comparison is with a world that has the data’s own persistence, not with zero.

The premium is an output: backing it out of one price

Every version of the equity risk premium answers the same equation. Write down the index price, write down a forecast of the cash its companies will pay, and solve for the discount rate that makes the two equal; subtract the bond yield, and the remainder is the premium that price implies. The price and the yield are observed. The forecast is not, and the premium inherits every assumption in it. That is why it cannot serve as an input: a valuation that assumes a premium has assumed the answer the price already gives. The single-company version of this backward solution is the reverse calculation in the discounted cash flow guide.

The yield gap reports a premium of -1.95 points, because it assumes rupee earnings never rise with prices. Let earnings rise with inflation and the same price implies 2.50 points on July 2026 inflation of 4.45 per cent (MoSPI, 12 August 2026, base 2024=100, provisional), or 2.05 on the target. The dividend route gives 2.54 at 8 per cent growth for ever and 6.59 at 12, a point of premium for each point of growth. The dividend route matches the earnings route on July’s inflation only at 7.96 per cent nominal dividend growth, a rate nobody observes. Payout, dividends over consolidated earnings, was 25.3 per cent, which is why the dividend form places so much of the value so far out.

The other popular estimate looks backward and averages what equities earned over bonds. Across the calendar years 2015 to 2025 the index beat the exchange’s 10-year benchmark government bond index by 5.34 points a year in log terms, with a standard deviation of 11.20. The 95 per cent interval on that average runs from -2.18 to +12.87 points. Eleven years cannot tell a premium of zero from one of twelve, so a historical premium used as an input is a guess carried to two decimal places.

What the comparison is for

The gap is a description, not a forecast: a statement of how the market prices equities against bonds under inflation and growth assumptions it does not state. Used that way it has three honest jobs. It shows whether equities got cheaper or dearer against bonds on one consistent earnings basis. It shows the premium the market is paying under an inflation assumption you name. And it shows what a change in yields would have to be made of, real rate or inflation, before it matters to prices.

It cannot time the market on this record. Its level broke on 31 March 2021, its predictive fit has to beat a null with its own persistence, and the only sample on one definition gives the wrong sign. The bond side is a price as much as a yield, as the guide to bond price and yield shows; the policy rate corridor sets where the curve starts, and the 10-year point is one reading of a whole curve, as the yield curve as a price list sets out. Separating what a ratio measures from what it is assumed to measure, and testing a claim against a null that keeps the data’s own persistence, is method rather than a figure to memorise, and it is how valuation is taught here.