RBI Bulletin Nov 2025

 The key highlights from the Reserve Bank of India Bulletin, November 2025, cover speeches on regulatory philosophy, technology integration, governance, economic resilience, central bank accounting, and several analytical articles providing deep dives into the Indian economy.

---

Key Headlines

Speeches:

• Regulation by RBI: Some Reflections by Shri Sanjay Malhotra
• The Evolving Facets of Regulations by Shri Sanjay Malhotra
• Transformational Technologies and Banking: Key Issues by Shri T Rabi Sankar
• Where Governance Intent is Strong, Regulatory Gaps and Overlaps Fade by Shri Swaminathan J
• Policy Frameworks for Economic Resilience: The Case of Emerging Markets and India by Dr. Poonam Gupta
• Central Bank Accounting Practices: The Reserve Bank of India and Global Approaches by Shri Shirish Chandra Murmu

Articles:

• State of the Economy
• ‘Making the Horizons Meet’: A Heterodox Approach for Short-Term Inflation Forecasting
• Multivariate Core Trend Inflation: A New Measure of Core Inflation
• Nowcasting GDP in India: A New Approach
• Seasonality in Key Economic Indicators of India
• Current Statistics
• Recent Publications

---

Summary of Key Highlights

I. Regulatory Philosophy and Governance

The Governor, Shri Sanjay Malhotra, detailed the philosophy and principles guiding RBI's regulation, emphasizing that the foremost priority and key objective is ensuring financial stability. Other objectives include ensuring the safety and soundness of financial operations through prudential requirements, consumer protection, assistance in law enforcement, and achieving broad socio-economic goals.

The key principles of regulation include a preference for principle-based formulation (often resulting in a hybrid of rules and principles, such as the proposed expected credit loss framework),, applying proportionality (customized, risk-sensitive regulations based on the scale and complexity of institutions),, relying on consultation for transparency, adopting an evidence and data-based approach, and committing to regular review (mandated every 5–7 years),.

Discussion of the evolving facets of regulation highlighted the strength of the banking sector, citing the rise in the Capital to Risk-Weighted Assets Ratio (CRAR) from 13.5% in March 2015 to 17.5% in March 2025 and a reduction in Gross Non-Performing Assets (GNPA) from 11.2% in March 2018 to 2.3% in March 2025,. Recent regulatory adjustments, such as revising Capital Market Exposure limits and withdrawing the Specified Borrower Framework, reflect the increased maturity and resilience of the banking system,.

Deputy Governor Shri Swaminathan J focused on governance intent, noting that regulatory gaps and overlaps tend to fade when intent is strong. He laid out principles for regulators, including balancing entity-based and activity-based regulation (regulating the activity wherever it occurs) and applying proportionality, demanding that rules be scaled according to risk and complexity,.

II. Technology and Banking Challenges

The speech on transformational technologies by Deputy Governor Shri T Rabi Sankar highlighted India's unique and successful model of leveraging Digital Public Infrastructures (DPIs) such as UPI, developed primarily through public sector initiatives and priced like public goods,. This infrastructure provided a foundational layer for private fintech firms to innovate,.

However, new technologies pose fundamental challenges to banks. The advent of digital currencies potentially offers an alternative to bank-created money, and blockchain technology could undermine the traditional role of banks as trusted intermediaries required to authenticate the payment leg of financial transactions,. To compete, banks must modernize their core systems, adopt a platform orientation, and cultivate a culture of innovation,.

III. Economic Resilience and Outlook (State of the Economy & Policy)

The Indian economy showed signs of a further pick up in momentum despite global headwinds, driven by strong urban and rural demand and support from Goods and Services Tax (GST) rate reductions and festive spending,.

Headline Consumer Price Index (CPI) inflation declined to an all-time low of 0.3 per cent in October 2025,. This sharp moderation was attributed largely to deepening deflation in food prices and the impact of GST rate cuts,.

The merchandise trade deficit widened to an all-time high in October 2025 due to a contraction in exports and a surge in imports, particularly gold and silver imports related to festive demand,. Financial conditions remained benign, with the total flow of financial resources to the commercial sector increasing significantly, primarily driven by non-bank sources,.

Dr. Poonam Gupta noted that India’s improved policy frameworks have resulted in economic and financial resilience. India has managed its external account effectively, maintained a low and stable external debt (18.9% of GDP as of end-June 2025),, and achieved successful outcomes with the Flexible Inflation Targeting framework. The near-term growth outlook is promising, with the growth forecast for FY2025-26 revised upwards to 6.8 per cent, and CPI inflation projected to be 2.6 per cent for the full year,.

IV. Analytical Contributions

• Short-Term Inflation Forecasting: A new hybrid Short-Term Forecasting Model (STFM) was introduced, integrating high-frequency nowcasts with structural models. This model shows improved predictive accuracy in the near-term by identifying the path of convergence from current data (nowcasts) to long-term benchmark forecasts, factoring in macro-linkages (e.g., a 10% exchange rate depreciation leads to approximately 80 basis points increase in inflation over 12 months),,.

• Core Inflation Measure: The Multivariate Core Trend (MCT) inflation measure was developed for India. This Bayesian model identifies underlying inflationary pressures by accounting for time-varying sectoral persistence and incorporating second-round effects from volatile components like food and fuel,. MCT inflation is found to be smoother and less volatile than conventional core CPI and offers superior predictive accuracy for core inflation over medium and long horizons,,.

• GDP Nowcasting: A novel two-step approach called the Two-Stage Maximum Information Model (TSMIM) was proposed for nowcasting GDP. This method uses the information often discarded by conventional models by creating a Secondary Composite Index (SCI) from the residuals of high-frequency indicators,. The empirical exercise demonstrated that incorporating this secondary information significantly improved the accuracy of nowcasting GDP and GVA compared to the benchmark Dynamic Factor Model (DFM),,.

• Seasonality: An analysis of economic indicators revealed stable, albeit often pronounced, seasonal patterns. Notable high seasonal variations were found in demand deposits of Scheduled Commercial Banks (SCBs), vegetable prices (particularly tomato, onion, and potato),, and merchandise exports. Quarterly analysis showed that both Real GDP and Gross Value Added (GVA) typically peak in the January-March quarter (Q4) and that GVA of agriculture experiences the highest seasonal variations among supply-side components,,.

Nowcasting GDP in India: A New Approach

by Indrajit Roy and K. M. Neelima The authors are from the Monetary Policy Department, Reserve Bank of India. The authors are thankful for the comments from anonymous reviewers and Shri Dipankar Biswas which have significantly enhanced the quality of the article. The authors thank Neha Dahale for data support. The views expressed in this article are those of the authors and do not represent the views of the Reserve Bank of India.

Nowcasting has become a useful tool for policymakers especially for macroeconomic variables like GDP for which data are released with considerable lags. A novel two-step approach for nowcasting India’s GDP is proposed using twenty-two high frequency indicators. The first factor is obtained based on the strength of each indicator in relation to GDP and a second factor is built from the residual information of the indicators which are otherwise generally discarded. The empirical exercise reveals that incorporating secondary information from residuals greatly improves accuracy of nowcasting GDP.

Introduction Following overlapping shocks—COVID-19 pandemic, multiple and prolonged geopolitical conflicts, surge in inflation across the globe—policy makers, especially central banks, had their task cut out to prop up the economy while managing the ramifications emanating from the different shocks. With new challenges emerging often, associated uncertainty has made it difficult to assess the current and future outlook of the economy. For policymakers, having a clear picture of the underlying state of the economy is critical for undertaking appropriate policy responses. Gross domestic product (GDP) can be considered as the most authoritative measure of economic activity. The national accounts provide a comprehensive view of the economy but are released with considerable lags. Therefore, forecasting macroeconomic indicators, especially GDP, is a necessity for optimal policy response and reliable short-term forecasts are generally high in demand when the economic environment is uncertain.

Central banks and other policy makers track certain high frequency indicators to gauge the underlying state of the economy. However, divergence among various indicators make it difficult to accurately assess the extant condition of the economy. Essentially, separating meaningful information from noise is a humongous task and several techniques ranging from detecting business cycle turning points and constructing indexes of economic activity to forecasting comprehensive macroeconomic measures of the state of the economy with formal models and judgment have been applied to tackle it. Summarising the information content available from different indicators using modelling techniques in recent periods into a composite index, or nowcasting, was thus developed to respond to the policymakers’ need for a reliable indicator in advance of the release of the relevant macroeconomic variable. Hence, an important feature of nowcasting is the extraction of all the available information from a large information set and it provides an early estimate of the reference series before it is published.

Most of the nowcast models generally reduces dimensions of large number of selected indicators into few factors. This can be done by dynamic factor modelling (DFM) or even using weighted average of indicators where weights are correlation of indicators with the target. High frequency indicators are generally selected for nowcasting based on convenience and timely availability of the indicators. As a result, for nowcasting a composite target, for example, GDP or gross value added (GVA), a sub-sector of the target may be over-represented by inclusion of greater number of selected proxy indicators vis-à-vis other sectors. Thus, factor modelling, which essentially reduces the dimensionality of large number of indicators and produces few common factors, is influenced by indicator selection bias. This occurs as the chosen factor may identify the co-movement of the selected indicators very well, while not necessarily capturing the relationship with the target series truly. It may also be the case that in DFM, information contents of the selected indicators are not completely utilised as only the first few latent factors are chosen based on eigenvalues, and other factors with lower eigen values are ignored which, in turn, may be having strong correlation with the reference series but are not strongly correlated with majority of other indicators in the information set.

In this article, an attempt has been made to nowcast India’s GDP using a two-step approach (two-stage maximum information model - TSMIM). In the first stage a primary composite index (PCI) is computed, by linearly combining these indicators based on the strength of their association (correlation) with the targeted series. PCI may also be computed using DFM. In the second stage, to further extract relevant information from the indicators beyond what is already extracted and aggregated in PCI, each indicator is regressed on the PCI and the corresponding residuals are estimated which are, in turn, aggregated based on their association with the target series to form a secondary composite index (SCI). PCI and SCI are then jointly used to nowcast GDP.

The novelty of our new approach lies in a) giving more weightage to those indicators that are correlated with the target series rather than co-movement of the indicators among themselves, and b) maximising information by utilising information which are discarded by the conventional modelling process by creating a secondary information source based on residuals.

To compare the nowcasting performance of TSMIM, following four models are considered: (a) TSMIM-1 with PCI as weighted average of chosen indicators and SCI as weighted average of residuals extracted from these indicators which are not part of PCI, (b) Using only PCI as weighted average of chosen indicators and no SCI, (c) TSMIM-2 with DFM based factor as the PCI and SCI as weighted average of residuals extracted from the indicators which are not part of DFM, and (d) Benchmark DFM. For ascertaining the efficiency of nowcasting exercise, out-of-sample prediction method is undertaken for the period Q4:2022-23-Q4:2024-25. This empirical exercise reveals a relatively improved performance of the new framework models (TSMIM-1 and TSMIM-2) when compared with one-step models (models b and d). We further check for robustness of the approach on real GVA growth and find that the new models demonstrate improved performance over DFM for GVA as well. Rest of the paper is organised as follows: brief literature review is undertaken in section II and section III elaborates the methodology and data used. Section IV discusses the results and the evaluation of the new model and section V concludes.

II. Literature Review There are many nowcasting techniques available in the literature. Among these techniques, principal component analysis/dynamic factor models (PCA/DFM) to nowcast low frequency macroeconomic variable such as GDP is popular. PCA, which is the core of the DFM model for nowcasting, transforms original information set into uncorrelated factors, which are the weighted linear combination of constituent indicators. Thus, it resolves both the curse of dimensionality issue as well as the multicollinearity issues.

A more prevalent approach of nowcasting currently is DFM. This method is widely used in summarising co-moving indicators, that cover the broad spectrum of economic activities in an economy, as latent factor(s) separated from idiosyncratic and measurement errors and can be interpreted as underlying state of the economy. In the model, a state-space framework is followed which involves a measurement equation linking the vector of observed indicators to a vector of unobserved state variables and a transition equation which specifies the dynamics of the unobserved state variables.

Stock and Watson (1989) pioneered the use of factor models for construction of business cycle indexes. Giannone et al. (2008), introduced factor models for nowcasting in a mixed frequency setup and the nowcasting model was developed using monthly data on a large set of high-frequency indicators. Many central banks have developed nowcasting methods to get a fair idea about how the economy is performing in a given quarter much before the official data release. The nowcasting model of the Federal Reserve Bank of New York uses a dynamic factor model that generates estimates of current quarter GDP growth at a weekly frequency. On the other hand, The Federal Reserve Bank of Atlanta’s GDPNow model is a nowcasting model that uses a bridge equation approach that relates GDP subcomponents to monthly source data with factor model and Bayesian vector autoregression techniques.

In India too, several studies have been undertaken on nowcasting GDP as GDP data are released by the National Statistics Office (NSO) with a lag of two months. Bhattacharya et al. (2011) finds that a small set of pre-selected key monthly indicators, serving as proxies for the various sub-sectors of the economy perform satisfactorily in predicting current quarter growth. Several studies in India use DFM for nowcasting. However, Matheson (2011) finds that forecasting performance of DFM for Australia and India was not on par with other countries which may be attributable to large data revisions of indicators in these countries. Bayesian vector autoregression and machine learning approaches are also used for nowcasting.

In DFM, the factors are linear combination of constituent indicators. However, as only first few components are chosen (with eigenvalue more than 1), and other factors with lower eigen values are ignored, there is an inherent loss of information since they might have good association with the reference series. It is likely that the factors with lower eigen value are those factors which are dominated (with higher loading/share) by indicators which may have strong correlation with the reference series but are not strongly correlated with majority of other indicators in the information set. As a result, there may be factors, which possess relevant information to explain variation in the reference series, but is discarded due to low eigen value which result in suboptimal performance in explaining the reference series. Our paper focuses on this aspect by obtaining additional information set from residuals which may help improve the nowcasting technique.

III. Methodology and Data

III.1 Two-stage Maximum Information Model (TSMIM) Following Roy and Narayanan (2018), we use a two-step process to construct a two-stage maximum information model (TSMIM). When the target series is of quarterly frequency and indicators are of monthly frequency, the indicators are averaged over three months to get quarterly series. If data for a variable is unavailable for a particular month, it is forecasted using ARIMA. These quarterly indicators are then standardised by subtracting their mean and dividing by standard deviation.

In the first step, we calculate correlation coefficient ($\rho_i$) of each indicator with the target series. Then primary composite index (PCI) is defined as follows:

$$ PCI_t = \sum_i x_{it} \cdot w'_i \cdot D_i \text{ where } w'_i = \frac{\rho_i}{\sum_i D_i \cdot \rho_i} \text{ and } D_i = 1 \text{ if } i^{th} \text{ indicator is significantly associated with } y_t \text{ and } D_i = 0, \text{ otherwise } $$

Essentially, PCI is a certain linear combination of selected indicators, and it contains common information of interest pertaining to association of these selected indicators with the reference series.

In the second step, we look for additional information in the information set beyond PCI. Certain components of the reference series may be under-represented or over-represented by the indicators selected, thus PCI may be influenced by this biased selection. We start with regressing each of the indicator on the PCI and estimating the corresponding residuals as follows:

$$ \text{Let } x_{it} = \eta_i + \delta_i \cdot PCI_t + \zeta_{it} \text{ where } \zeta_{it} \sim i.i.d. N (0, \psi_i^2), \eta_i \text{ and } \delta_i \text{ are unknown coefficients} $$

Let $\hat{\zeta}{it}$ where $i = 1, 2, \ldots, n$ be the residuals estimated from equation (3) corresponding to $i^{th}$ indicator and these residuals form the constituents of the additional information set. Let $r_i$ is the correlation coefficient of $\hat{\zeta}{it}$ with $y_t$. Secondary composite index (SCI) is computed as follows:

$$ SCI_t = \sum_{i=1}^k \hat{\zeta}_{it} \cdot w''_i \cdot E_i \text{ where } w''i = \frac{r_i}{\sum_i E_i \cdot r_i} \text{ and } E_i = 1 \text{ if } \hat{\zeta}{it} \text{ is significantly associated with } y_t \text{ in equation (3), and } E_i = 0, \text{ otherwise} $$

Notably, this additional information set is built from the residual information of the $n$-indicators which are not part of PCI. Therefore, PCI and SCI are two composite indices of coincident indicators for the reference series constructed out of many indicators and are linear combination of indicators. Also, by construction, $PCI_t$ and $SCI_t$ are uncorrelated, therefore, these can be used together to explain $Y_t$ without any multicollinearity issue. Nowcasted value of the reference series can be obtained as follows:

$$ Y_{t+1} = \eta_0 + \eta_1 \cdot PCI_{t+1} + \eta_2 \cdot SCI_{t+1} + \epsilon_{t+1} $$

The estimated value of $Y_{t+1}$ in equation (5) is a weighted average of PCI and SCI. Together, PCI and SCI can explain the reference series much better than PCI alone.

III.2 Data For nowcasting a reference series, a set of indicators is chosen based on the strength of their connection or relationship (correlation coefficients, visual inspection from scatter plot) with the reference series. We initially followed Kumar (2020) for variable selection as the 27 indicators in the model were based on whether they are tracked by NSO, their correlation with GDP, and availability of time series data. However, we find that five variables viz., US PMI, non-food credit, tractor sales, CPI excluding food and beverage and money supply were not significantly correlated with GDP and were dropped from the model. We use a set of 22 indicators which are significantly correlated with GDP. Broadly, these indicators cover major segments of domestic activity- directly or indirectly. The four blocks of data are a) industry; b) services, c) global and d) miscellaneous. The data includes hard data on economic activity for example, index of industrial production, automobile sales, port cargo traffic, domestic air cargo traffic etc., b) surveys representing economic activity like PMI, c) trade such as non-oil exports, non-oil imports, Baltic Dry Index etc., d) employment conditions as captured by JobSpeak Index and e) global conditions as captured by OECD composite leading indicator, US payroll data and US industrial production. All indicators are in year-on-year terms. The nowcasting exercise is undertaken separately for each quarter for the period Q4:2022-23 – Q4:2024-25 using data spanning Q1:2011-12 – Q4:2024-25. The data were winsorised at 0 and 99.25 levels to adjust for outliers caused by COVID-19 pandemic.

IV. Results: Nowcasting of India’s GDP and Model Evaluation The coefficients—PCI, SCI and dynamic factor (DF)— are significant suggesting that these factors can explain real GDP growth. Further, the different models for nowcasting GDP show that the in-sample fit of (model a) TSMIM-1, and (model c) TSMIM-2 are better than that of using benchmark DFM model (model d), and using only PCI (model b) on the basis of R-square. Using a single factor - DF or PCI - explains around 85 per cent of the variability of GDP while using two factors together explain 94 per cent of the variability of GDP suggestive of superior performance of using secondary factor in both cases. Nowcasted GDP using TSMIM-1 was found to be closely following the target variable.

The RMSE of SPF and model (b) were found to be the highest among all the sets of nowcasts. RMSE of DFM is the next highest primarily on account of the inability of the model to predict the growth in Q1:2023-24. TSMIM-1, followed by TSMIM-2 had the lowest RMSE in the period under study crucially underpinning the role of secondary information in improving nowcast accuracy even while excluding Q1:2023-24 from calculation of RMSE. The models using SCI have performed consistently better than the models using only single factor in most quarters in nowcasting GDP (closer to the actual estimates) from the information content available through the same set of twenty-two indicators. During the period of high growth in 2023-24, TSMIM-1 and TSMIM-2 nowcasts were closer to the latest GDP estimates than those generated by using only DF and PCI in single-step models underpinning higher accuracy achieved due to the use of secondary factor.

IV.1.a Robustness Checks The same exercise was undertaken for nowcasting real gross value added (GVA) growth at basic prices for the same period. The regression estimates for Q4:2024-25 for GVA show that the coefficients of the variables of interest are significant in all models. As in the case of GDP, the model fit of TSMIM-1, and TSMIM-2 are better than that of using only DF in the case of GVA as well. RMSE of nowcasts for real GVA growth generated using TSMIM-1 and TSMIM-2 are the lowest among all models, which improves further if Q1:2023-24 is excluded from the sample. Incidentally, RMSE of TSMIM for nowcasting GVA is lower than that of nowcasting GDP reflective of indicators capturing economic activity from supply side. Visually, TSMIM-based nowcasts track GVA more closely than DFM based nowcasts, especially in the latest quarters.

V. Conclusion Many important low-frequency macro-economic indicators such as GDP are subject to publication delays or lag. However, there are many high frequency coincident indicators which are correlated with the targeted macroeconomic indicator that are available at much shorter time lags. Monitoring all these coincident indicators and revising the assessment about the reference series is a difficult task for the policy makers. Combining all these coincident indicators into a composite index for nowcasting was thus developed to meet the need for reliable indication in advance of the release of the relevant macroeconomic indicator.

This article uses a new framework (TSMIM), to extract maximum information relevant to nowcast the reference series. Coincident indicators are combined into a weighted composite index (PCI), which generally tracks the reference series well. However, residual information may contain some more information for the reference series which are not completely captured by PCI. Therefore, the new approach further extracts information which is not part of the already calculated composite index i.e., PCI, and has potential to be related to the reference series. These secondary indicators derived from the primary set of indicators are then again combined into a SCI with suitable weights.

This article shows a relatively improved performance (both in-sample and out of sample) of the new framework (TSMIM) when compared with the baseline pure dynamic factor-based modelling (DFM) based nowcasting. TSMIM-based nowcasts track GDP and GVA more closely than pure DFM based nowcasts, when the nowcasts generated for each vintage vis-à-vis the actual official data are compared. Moreover, TSMIM framework can accommodate DFM model and can reduce forecast errors further. DFM can be used in the first stage of TSMIM and in the second stage, residual information are extracted from the coincident indicators which are not part of the dynamic factor obtained from DFM and these residuals are combined into a second factor with a suitable weights (such as correlation with the reference series). DF and second factor thus obtained produced improved nowcast of the reference series than nowcasts generated by benchmark DFM alone. Incorporating secondary information from residuals greatly improves the accuracy of nowcasting and therefore, TSMIM is an important addition to the nowcasting toolkit.