## wpiea2025126-print-pdf ## Source details **Canonical URL:** [wpiea2025126-print-pdf](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025126-print-pdf.pdf) ## Other formats - [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025126-print-pdf.pdf.md) - [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025126-print-pdf.pdf.json) --- ### Setting and descriptive evidence - Provides institutional background on UPI and interoperability, data description, and descriptive evidence that users value interoperability. - Presents two new stylized facts suggesting interoperability supported adoption of digital payments in India. - Motivation for a theoretical model (Section 3) and empirical strategy (Section 4) to quantify causal impacts of interoperability. - Institutional context (UPI and interoperability): - UPI is an instant payments platform built on top of the Immediate Payment Service (IMPS) infrastructure. NPCI developed UPI and also provides IMPS payment rails; NPCI is regulated by the Reserve Bank of India. - Any UPI app user can initiate payments from accounts at any participating bank to accounts at other participating banks and receive notifications of payments received. - UPI apps provided by banks and non-bank third-party application providers (TPAPs). - Bank apps: bank provides front end and executes transactions via IMPS. - TPAP apps: TPAP front end partners with a payment service provider (PSP) bank connected to IMPS. - UPI enabled transactions previously not possible by increasing interoperability across banks and closed-loop e-money providers. - Settlement: end-user settlement is immediate; settlement among financial institutions managed through deferred net settlement with ten daily cycles. - Interoperability effect on users: - End users can choose their UPI app without affecting location of deposits or set of other UPI users they can transact with. - Contrasts with closed-loop apps where payer and payee must use the same app, creating provider-level network effects. - Under UPI, network effects operate primarily at the level of the overall UPI ecosystem. - Growth and enabling background: - Several hundred participating apps and banks. - Transaction volume grown to more than 23 billion transactions per month. - UPI dominates other forms of electronic retail payments in India and proxies for cash beginning to decline. - India now makes more fast payments than any other country. - Enabling policies: Pradhan Mantri Jan Dhan Yojana (JDY); Aadhaar biometric ID scheme; mobile data cost fell by roughly 96 percent during the mid-2010s. ### Data sources (four key sources) - Novel universe-level UPI transactions data: - Monthly totals of value and volume, split by interaction of the payer’s app and the payer’s bank branch, from April 2016 through September 2024. - For the largest three UPI apps, plus BHIM and a consolidated “Other Apps” category: full matrix of payer and payee app choices (monthly totals of value and volume split by payer’s bank branch, payer’s app, and payee’s app). - Transactions disaggregated into peer-to-peer (P2P) and peer-to-merchant (P2M). - Unique users proxied by unique phone numbers per month at the IFSC level. - Data from a major Indian fintech firm (“the incumbent”): - Firm pre-dated UPI; offered a closed-loop wallet before integrating with UPI. - Monthly, district-level totals of value, volume and unique users, split by P2P and P2M. - NPCI data on cross-bank ATM cash withdrawals: - April 2016 to December 2023, split by bank and pincode; aggregated to district level. Used as a proxy for cash usage. - Consumer Pyramids Household Survey (CPHS) from CMIE: - Panel survey covering most Indian districts with household borrowing data. - District-level demographics from the 2011 census; bank branches identified by IFSC; Appendix B cross-checks location proxy. ### New stylized facts (empirical patterns) - Fact 1: Cross-app transactions are an important element of activity on UPI. - Share of UPI transactions where payer and payee use different UPI apps consistently tops 40% on both value and volume. - Cross-app transactions were especially large in early years; persistence suggests users value ability to pay users of different apps. - Fact 2: Users’ post-demonetization choices indicate preference for interoperable over closed-loop payments. - Following demonetization, both UPI and the incumbent closed-loop platform saw sharp increases in November and December 2016. - As cash availability returned in early 2017: - Adoption of the non-interoperable incumbent was flat between March and October 2017. - UPI grew roughly three-fold between March and October 2017. - Cross-app payments also rose rapidly. - Patterns consistent with users preferring interoperable platforms when forced to try digital payments. - Limitation: - Descriptive patterns do not identify causal counterfactual without interoperability; paper proceeds to theoretical and empirical analysis. ### Model: setup and key mechanisms - Objective: stylized static model where users choose between two digital platforms (a and b) and outside option C (cash); shows interoperability can increase digital payments usage by unifying fragmented networks, with larger increases in more fragmented districts. - Environment: - Districts d ∈ {1,...,D}; each district contains a closed unit square of users making within-district payments. - Users distributed (x,y) ∼ U([0,1] × [0,1]); x and y capture intrinsic preferences. - Platforms a, b, and cash C are mutually exclusive choices. - Digital payment convenience utilities (heterogeneous valuations): - u_a,d,x,y = { 1 + κ N*_d,a if x ≤ x̂_d ; 0 if x > x̂_d } - u_b,d,x,y = { 0 if x ≤ x̂_d ; 1 + κ N*_d,b if x > x̂_d } - N*_d,i is number of digital payments users accessible through platform i in district d. - κ > 0; x̂_d ∈ (0, 1/2) denotes share of “a-type” users; higher x̂_d indicates more pre-interoperability fragmentation. - Cash convenience: u_C,d,x,y = γ y, with γ > 1 + κ ensuring some users always choose cash. - Interoperability: when imposed, N*_d,i becomes N^D_d := N_d,a + N_d,b for both i ∈ {a,b}. ### Equilibrium analysis (3.2) - Users choose p ∈ {a,b,C} to maximize U_d,x,y; tie-breaking: indifferent users between cash and digital choose digital. - Benchmark: homogeneous valuations - Lemma 1: Only stable equilibria have all digital users pooled on one platform. - Total digital usage: N_D,Homog_d = ̄y = 1/(γ−κ). - Baseline: heterogeneous valuations and fragmentation - Lemma 2: - N_d,a = ˆx_d ˆy_d,a = ˆx_d γ−κˆx_d - N_d,b = (1−ˆx_d) ˆy_d,b = 1−ˆx_d γ−κ(1−ˆx_d) - Total digital usage: N_D,Baseline_d = ˆx_d γ−κˆx_d + 1−ˆx_d γ−κ(1−ˆx_d) - Note: a third equilibrium with equal usage exists but is unstable. - Interoperability equilibrium and network benefits - Lemma 3: - Under heterogeneous valuations, interoperability raises total digital payments to ̄y = 1/(γ−κ). - N_d,a = ˆx_d ̄y = ˆx_d γ−κ - N_d,b = (1−ˆx_d) ̄y = 1−ˆx_d γ−κ - Proposition 1: Interoperability resolves fragmentation, increasing total usage to the homogeneous-valuation level. - Proposition 2: Interoperability increases usage of both platforms relative to baseline without interoperability. - Comparative statics: fragmentation and gains - Proposition 3: More pre-interoperability fragmentation yields larger increases in total digital usage from interoperability. - Proposition 4: Impact on usage per platform depends on initial fragmentation; marginal effects can be non-monotonic for a given platform as ˆx_d varies. - Mechanism: aggregate gains increase in ˆx_d over 0 < ˆx_d < 1/2 with diminishing slope; as ˆx_d → 1/2 marginal changes have no impact on aggregate gain. - Empirical predictions (3.3): - Observable fragmentation measure F_d = N_d,a,−1 / N_D_d,−1 (pre-interoperability share of smaller platform). - Prediction 1: Introducing interoperability increases total digital usage by more in districts with higher pre-interoperability fragmentation: ∂∆N_D_d / ∂F_d > 0. - Prediction 2: When pre-interoperability fragmentation is relatively low, interoperability increases usage on each platform more in districts with higher fragmentation: ∂∆N_d,a / ∂F_d > 0 and ∂∆N_d,b / ∂F_d > 0. - Appendix D addresses external shocks ω and unobservability of ˆx_d. - Modeling choices discussion (3.4): - Focus on adoption (number of users) rather than transactions; empirical decomposition covers (i) average value per transaction, (ii) number of transactions per user, (iii) number of users per capita, with margin (iii) accounting for majority of variation. - Model abstracts from platform provider decisions; interoperability exogenous. - Cross-district payments (Appendix F) would attenuate estimated impact, implying empirical estimates are likely a lower bound. - Preference boundary ˆx_d assumed exogenous; removing that margin would not change main empirical predictions if switchers are few. ### Estimation strategy (4.1) - Empirical setting: integration with UPI of a major incumbent fintech firm that previously offered closed-loop payments; incumbent had transaction value similar to UPI in the month prior but UPI was growing faster. - Geographic scope: UPI total transaction value higher than incumbent’s in 97% of districts in month before integration. - Fragmentation measure Fd: - Fd := Total value of transactions per capita on the smaller platform in district d in t−1 / Total value of transactions per capita across both platforms in district d in t−1. - Interpretation: values closer to 50% imply higher fragmentation; median(Fd) = 7.4%. - Construct dummy F+d = 1 for districts with Fd above the median. - Baseline regression (district-month level): - ydt = αd + αst + β(F+d × 1{t≥t0}) + βZ (Zyd × 1{t≥t0}) + edt - αd and αst are district and state-time fixed effects; Zyd = overall value of transactions per capita across both platforms in t−1. Standard errors clustered at district level. Top 1% of ydt winsorized. - Event-study: - ydt = αd + αst + Στ≠t−1 βτ (F+d × 1{t=τ}) + Στ≠t−1 ατ (Zyd × 1{t=τ}) + vdt - Outcomes and exclusions: - Primary outcome: total peer-to-merchant (P2M) transaction value across both platforms per capita (P2P excluded to focus on within-district payments). - Identification concerns and mitigations: - Anticipation: RBI directive mandated wallets make wallets interoperable through UPI immediately prior to integration; widespread accurate anticipation deemed implausible. Event-study shows no differential pre-trends; any anticipation would bias β downward. - Parallel trends and confounders mitigations: 1. State-time fixed effects. 2. Control for pre-integration digital payments usage Zyd. 3. No significant pre-integration trend differences from event-study. 4. Robust to matched sample of low-Fd districts. 5. Robust when instrumenting Fd with geographic decisions by the incumbent taken more than a year prior to integration. - Auxiliary design choices: - Estimation window: six months before to one year after integration. - Clustering: standard errors clustered at district level. - Winsorization: top 1% of ydt winsorized. - Matching: Mahalanobis distance matching on log population with replacement; retained 500 districts; discarded 54 low-Fd districts never matched and 19 high-Fd districts with populations outside low-Fd range. - Instrumental variable: H d = negative distance in kilometers from district centroid to nearest “hub” centroid; eight hub districts with at least 1000 merchants adopted by September 2016. Pre-2016 merchant counts: 92% of districts had fewer than 100 merchants, 85% had fewer than 10, and 42% had zero. ### Aggregate national impact (5.1) - Object of interest: ∆I_ND, total national change in digital payments usage from integrating fragmented networks. - Identification challenge: “missing intercept” — unknown absolute impact in less fragmented districts. - Theoretical result (Proposition 5): ∆I_ND = Σ_d (∆N^D_d − ∆N^D_d0), where ∆N^D_d0 is change in a district already fully unified prior to interoperability. - Empirical proxy for “fully unified ex ante” districts: first decile of F_d distribution (median F_d of 1% in that decile). - Relative-decile baseline specification: - y_dt = α_d + α_st + Σ_{n=2}^{10} β_n (F^n_d × 1{t≥t0}) + β_Z (Z^y_d × 1{t≥t0}) + e_dt - β̂_n estimates monthly average increase in total P2M transaction value per capita for decile n relative to first-decile districts. - Aggregation to national impact: - ∆I_y = Σ_d Population_d / Σ_d Population_d × Σ_{n=2}^{10} β̂_n × F^n_d - Procedure: compute β̂_n × F^n_d for each district and decile, aggregate with population weights, renormalize by national population. - Key numerical results: - Estimated population-weighted national average monthly increase in total P2M transaction value per capita attributable to interoperability: - ∆I_y = 10.2 Rupees per capita per month during the first year after integration. - Integration raised average national P2M transaction value per capita per month by 59% relative to the model-implied no-integration counterfactual. - Interpretation and scope: - Aggregate result captures total reduced-form effect of network integration, including direct cross-platform transactions and induced additional activity from expanded accessible network. - Given UPI integrated additional pre-existing networks (none as large as incumbent), findings suggest interoperability played a significant role in UPI’s take-off. - Caveat: aggregation inherits model assumptions; estimates may under-state true impact (e.g., cross-district transaction gains netted out). ### Downstream lending spillovers and supply-side considerations - Downstream lending (summary of Section 5.2): - Household-level specification: - y_ht = α_h + α_st + β (F^+_d × 1{t≥t0}) + β_Z (Z^y_d × 1{t≥t0}) + e_ht - Outcome: probability of borrowing from non-banks (NBFCs). - Empirical findings: - In above-median fragmentation districts, probability of borrowing increases by 1.1 percentage points relative to below-median districts after integration. - Larger increases for entrepreneurs and hawkers. - Similar results for households with no prior borrowing. - Interpretation: interoperability-induced payments activity appears to reduce credit frictions and enable more borrowing. - Long-term supply-side considerations (summary of Section 5.3): - Ambiguity: interoperability can generate coordination complementarities or free-rider under-investment. - Evidence of provider responses recreating proprietary within-app effects: - Proprietary QR codes (prohibited later), branded interoperable QR codes suggesting app-specific readability, referral schemes incentivizing within-app transactions. - Observed patterns: - Roughly half of UPI transactions remain within-app; cross-app to within-app ratio broadly stable in recent years. - Increasing regional concentration in user bases of the two largest apps. - Policy implication: - Maintaining a truly interoperable system may require continued regulatory and policy efforts to counter provider strategies that re-introduce lock-in and proprietary network effects. - Recommendation for future work: - Fuller assessment of long-term supply-side impacts and welfare implications as more provider-level data become available. ### Key empirical estimates and tables (highlights) - Table 1 (district-level regressions; six months before to one year after): - F+d × 1{t≥t0} coefficients: - Column (1) Total/pop: 8.342 ∗∗∗ (t-stat 4.67), N = 10,868. - Column (2) Incumbent/pop: 4.925 ∗∗∗ (t-stat 6.03), N = 10,868. - Column (3) Other/pop: 2.971 ∗∗∗ (t-stat 3.70), N = 10,868. - Column (4) (Inc↔Oth)/pop: 0.208 ∗∗∗ (t-stat 4.13), N = 7,436. - Mean ydt (F+d = 1,t = t−1): 9.018, 7.089, 1.928 (Columns 1–3). - Mean ydt (F+d = 0,t≥ t0): 6.687, 1.450, 5.294, 0.189 (Columns 1–4). - Table 2 (matching on log population): - F+d × 1{t≥t0} coefficients: - Column (1): 7.002 ∗∗∗ (t-stat 4.82), N = 9,500. - Column (2): 4.524 ∗∗∗ (t-stat 6.40), N = 9,500. - Column (3): 2.364 ∗∗∗ (t-stat 3.16), N = 9,500. - Column (4): 0.172 ∗∗∗ (t-stat 3.41), N = 6,500. - Table 3 (2SLS instrumenting with proximity to incumbent hubs): - F+d × 1{t≥t0} coefficients: - Column (1): 17.76 ∗∗∗ (t-stat 2.74); K-P F -Stat 20.44, N = 10,621. - Column (2): 10.01 ∗∗∗ (t-stat 2.97); K-P F -Stat 19.88, N = 10,621. - Column (3): 6.311 ∗∗ (t-stat 1.98); K-P F -Stat 21.43, N = 10,621. - Column (4): 0.642 ∗∗ (t-stat 2.37); K-P F -Stat 3.15, N = 7,267. - Table 4 (household-level NBFC borrowing response): - F+d × 1{t≥t0}: - Column (1) All: 0.0113 ∗∗ (t-stat 2.17), N = 898,412. - Column (2) Entrepreneurs: 0.0192 ∗∗ (t-stat 2.54), N = 54,161. - Column (3) Hawkers: 0.0136 ∗∗∗ (t-stat 3.00), N = 22,387. ### Robustness, appendices, and methodological notes - Estimation robustness: event-study shows no differential pre-trends; matched-sample and IV specifications yield similar qualitative results. - Appendices cover proofs, extended model with external shocks (Appendix D), multihoming (Appendix E), cross-district transactions (Appendix F), decomposition of value per capita change (Appendix G), additional robustness (Appendix H), and estimation of “Other-Other” cross-app transactions (Appendix I). - Location proxy cross-check: correlation coefficient 0.84 between NPCI-mapped PhonePe transactions and PhonePe Pulse (Figure B.1a). - Theoretical extension includes unanticipated shock ω with 0 < ω < 1 − 1/(γ−κ) (Appendix D). *Source: wpiea2025126-print-pdf - introduction of interoperability. Section 4 describes our empirical strategy and tests the model’s. Canonical URL: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025126-print-pdf.pdf* ### introduction of interoperability. Section 4 describes our empirical strategy and tests the model’s ### introduction of interoperability. Section 4 describes our empirical strategy and tests the model’s ### Setting and Descriptive Evidence — overview - Section provides background on the institutional context, describes the data, and presents descriptive evidence that users value interoperability. - Two new stylized facts are presented that suggest interoperability supported adoption of digital payments in India. - Motivates the need for a model (Section 3) and empirical strategy (Section 4) to quantify causal impacts of interoperability. ### Institutional context (UPI and interoperability) - UPI is an instant payments platform built on top of the Immediate Payment Service (IMPS) infrastructure. - NPCI developed UPI and also provides IMPS payment rails; NPCI is regulated by the Reserve Bank of India. - Users of any UPI app can initiate payments from their accounts at any participating bank to accounts at other participating banks, and receive notifications of payments received into their accounts. - UPI apps are provided by both banks and non-bank third-party application providers (TPAPs), typically fintech firms. - Bank apps: bank provides user-facing front end and executes transactions on the back end through IMPS. - TPAP apps: TPAP provides front end and partners with a payment service provider (PSP) bank connected to IMPS that executes the transaction. - UPI enabled transactions that were previously not possible: - prior options included (i) bank-to-bank transfers via IMPS, (ii) transfers between some bank accounts and electronic wallets issued by closed-loop e-money providers, (iii) transfers between electronic wallets hosted by the same closed-loop e-money provider; - UPI allowed TPAPs to interact with NPCI’s IMPS via a partner PSP, increasing interoperability across banks and closed-loop e-money providers. - Settlement: settlement for end users is immediate; settlement among financial institutions is managed through deferred net settlement with ten daily cycles. - Interoperability effect on users: - end users can choose their favorite UPI payments app without affecting location of their deposits (which remain in their bank) or the set of other UPI users with whom they can transact. - contrasts with closed-loop digital payment apps, where both payer and payee must use the same payment app, creating provider-level network effects. - under UPI, network effects operate primarily at the level of the overall UPI ecosystem. - UPI ecosystem growth and enabling background: - several hundred participating apps and banks (Figure 1 referenced). - transaction volume has grown exponentially to more than 23 billion transactions per month. - UPI now dominates other forms of electronic retail payments in India and proxies for cash beginning to decline (Figure 2 referenced). - India now makes more fast payments than any other country (Figure 3 referenced). - Facilitating investments and policies: - Pradhan Mantri Jan Dhan Yojana (JDY) opened hundreds of millions of new bank accounts. - Aadhaar biometric ID scheme provided unique verifiable digital identity aiding authentication and KYC. - cost of mobile data fell by roughly 96 percent during the mid-2010s, driven in part by entry of Reliance Jio (a new 4G-only network operator). ### Data — four key sources - Novel universe-level UPI transactions data: - monthly totals of value and volume, split by interaction of the payer’s app and the payer’s bank branch, from April 2016 through September 2024. - for the largest three UPI apps, plus BHIM and a consolidated “Other Apps” category, the full matrix of payer and payee app choices (monthly totals of value and volume split by payer’s bank branch, payer’s app, and payee’s app). - transaction totals are disaggregated into peer-to-peer (P2P) and peer-to-merchant (P2M) transactions. - number of unique users proxied by unique phone numbers per month at the IFSC level. - Data from a major Indian fintech firm (“the incumbent”): - firm pre-dated UPI, offered a popular closed-loop wallet before integrating with UPI. - monthly, district-level totals of value, volume and unique users, split by P2P and P2M. - NPCI data on cross-bank ATM cash withdrawals: - April 2016 to December 2023, split by bank and pincode; aggregated to district level. - used as a proxy for cash usage. - Consumer Pyramids Household Survey (CPHS) from CMIE: - panel survey covering most Indian districts, providing granular, high-frequency panel data on household borrowing. - District-level demographic and descriptive data from the 2011 census; all district-level data homogenized to consistent district boundaries corresponding with this census. - Bank branches are identified by Indian Financial System Codes (IFSC); IFSC used to extract bank name and pincode to attribute transactions to banks and districts. Appendix B cross-checks this location proxy. ### New stylized facts (empirical patterns) 1. Cross-app transactions are an important element of activity on UPI. - Figure 4 shows share of UPI transactions where payer and payee use different UPI apps. - On both value and volume, the share of cross-app transactions consistently tops 40%. - Cross-app transactions were especially large in the early years when the platform reached widespread adoption. - Persistence of cross-app transactions suggests preferences for payment apps vary across users and that users directly value and utilize ability to pay users of different apps. 2. Users’ post-demonetization choices indicate a preference for interoperable over closed-loop payments. - Following demonetization, both UPI and the incumbent closed-loop platform experienced sharp increases in transaction values in November and December 2016. - As cash availability returned to normal in early 2017: - growth in usage of the non-interoperable closed-loop incumbent plateaued (adoption of the non-interoperable alternative was flat between March and October 2017), - UPI grew roughly three-fold between March and October 2017, - cross-app payments also rose rapidly (Appendix Figure A.2 referenced). - These patterns are consistent with users preferring interoperable platforms when forced to try digital payments. - Limitations of stylized facts: - these descriptive patterns do not allow for causal or quantitative inference about the counterfactual without interoperability; the paper proceeds to theoretical and empirical analysis to address these questions. ### Model — setup and key mechanisms - Purpose: - stylized static model where users (households and firms) choose between two digital payments platforms (a and b) and an outside option C (cash). - model shows how interoperability can increase usage of digital payments by unifying otherwise fragmented networks, with larger increases in more fragmented districts. - empirical predictions derived in Section 3.3 and tested in Section 4; proofs in Appendix C. - Environment: - many districts d ∈ {1,...,D}; each district contains a closed unit square of users; each user makes a within-district payment. - three mutually exclusive payment methods: platform a, platform b, outside option C (cash). - users distributed uniformly along (x,y) ∼ U([0,1] × [0,1]); x and y capture intrinsic preferences over payment methods. - all users choose p_d,x,y ∈ {a,b,C} simultaneously. - model abstracts from dynamic early/late adoption considerations. - Preferences for digital payments: - convenience utility for user (x,y) in district d for platform i ∈ {a,b}: - u_a,d,x,y = { 1 + κ N*_d,a if x ≤ x̂_d ; 0 if x > x̂_d } - u_b,d,x,y = { 0 if x ≤ x̂_d ; 1 + κ N*_d,b if x > x̂_d } - N*_d,i is number of digital payments users accessible through platform i in district d. - in absence of interoperability, N*_d,i = N_d,i (number of users of platform i in district d). - κ > 0 summarizes intensity of network benefits generated by each accessible user. - boundary x̂_d ∈ (0, 1/2) denotes share of users in each district that perceive benefits from platform a vs platform b (share of “a-type” vs “b-type” users). - districts differ in x̂_d; lower bound x̂_d > 0 implies both types present in every district. - upper bound x̂_d < 1/2 imposed for expositional convenience so platform b ends up with more users pre-interoperability; higher x̂_d is a monotonic indicator of pre-interoperability network fragmentation. - Preferences for cash: - cash utility u_C,d,x,y = γ y. - N_d,C denotes number of cash users in district d. - cash convenience does not depend on other users’ choices in benchmark; heterogeneity in y captures idiosyncratic cash preference or adoption costs. - parameter restriction γ > 1 + κ ensures some users always choose cash. - Expectations and equilibrium: - users have rational expectations and model focuses on stable, rational equilibria in pure strategies. - equilibrium: {N_d,a, N_d,b, N_d,C} with N_d,a + N_d,b + N_d,C = 1 for all d; users’ expectations about totals are correct; small deviations revert to same equilibrium (stability). - Interoperability in the model: - when interoperability is imposed, each digital payments platform enables access to the combined user base of both: - N*_d,i becomes N^D_d := N_d,a + N_d,b for both i ∈ {a,b}. *Source: wpiea2025126-print-pdf - introduction of interoperability. Section 4 describes our empirical strategy and tests the model’s* ### 3.2 Equilibrium analysis ### 3.2 Equilibrium analysis ### Users’ decision problem - Users choose payment method p ∈ {a, b, C} to maximize U_d,x,y: - U_d,x,y = u_a_d,x,y if choosing digital payments platform a, - U_d,x,y = u_b_d,x,y if choosing digital payments platform b, - U_d,x,y = u_C_d,x,y if choosing cash. - Tie-breaking assumption: users indifferent between cash and a digital platform choose the digital platform. ### Benchmark: homogeneous valuations - Model specialization: replace equation (1) with u_i_d,x,y = 1 + κ N_d,i for i ∈ {a, b}. - Lemma 1 (Homogeneous platform valuations): - When users’ valuations of the two digital platforms are homogeneous, the only stable equilibria are those in which all digital payments users pool on one platform. - Total digital payments usage in these equilibria: N_D,Homog_d = ̄y = 1/(γ−κ). ### Baseline: heterogeneous valuations and fragmentation - With preferences following equation (1), exogenous differences lead to fragmentation between platforms. - Lemma 2 (Baseline equilibrium): - Usage of platform a: N_d,a = ˆx_d ˆy_d,a = ˆx_d γ−κˆx_d - Usage of platform b: N_d,b = (1−ˆx_d) ˆy_d,b = 1−ˆx_d γ−κ(1−ˆx_d) - Total digital payments usage: N_D,Baseline_d = ˆx_d γ−κˆx_d + 1−ˆx_d γ−κ(1−ˆx_d) - Note: a third equilibrium with equal usage of the two platforms exists but is unstable. ### Interoperability equilibrium and network benefits - Lemma 3 (Interoperability equilibrium): - Under heterogeneous valuations, interoperability raises total digital payments to the homogeneous-valuation level. - Usage of platform a: N_d,a = ˆx_d ̄y = ˆx_d γ−κ - Usage of platform b: N_d,b = (1−ˆx_d) ̄y = 1−ˆx_d γ−κ - Total digital payments usage: N_D,Interop_d = ̄y = 1/(γ−κ) - Proposition 1 (Interoperability and total digital payments): - Interoperability resolves market fragmentation, increasing total usage of digital payments to the level that would result in the absence of heterogeneous valuations. - Proposition 2 (Interoperability and usage per digital platform): - When users’ valuations are heterogeneous, interoperability increases usage of both platforms relative to the baseline without interoperability. ### Comparative statics: fragmentation and gains from interoperability - Proposition 3 (Impact on total digital payments by fragmentation level): - The more fragmented users are across platforms absent interoperability, the larger the increase in total digital payments usage unlocked by interoperability. - Intuition and illustration: - Districts with higher ˆx_d (more fragmentation) see larger increases in total digital payments under interoperability, converging to the same ̄y = 1/(γ−κ). - As ˆx_d increases from 0 toward 1/2, the aggregate gain from interoperability increases but with a declining gradient. - Proposition 4 (Impact on usage per platform by fragmentation level): - If the no-interoperability level of fragmentation is relatively low, a marginally higher fragmentation leads to a larger increase in usage of both platforms under interoperability. - If the no-interoperability level of fragmentation is relatively high, a marginally higher fragmentation can lead to either a larger or smaller increase in usage of a given platform under interoperability. - Mechanism: - Aggregate gains from interoperability are increasing in ˆx_d over 0 < ˆx_d < 1/2 but with diminishing slope; in the limit as ˆx_d → 1/2 a marginal change in ˆx_d has no impact on the aggregate gain. ### 3.3 Empirical predictions - Two empirical complications addressed in Appendix D: - External shocks ω occurring alongside interoperability that can affect adoption. - Unobservability of ˆx_d; introduce observable fragmentation measure F_d = N_d,a,−1 / N_D_d,−1, where N_d,a,−1 and N_D_d,−1 are observed usage of the smaller platform and of both platforms combined in the pre-interoperability equilibrium. - Using extensions accounting for ω and observable F_d, testable predictions (robust to external shocks if parallel trends hold): - Prediction 1 (Interoperability and total digital payments): - Introducing interoperability increases total usage of digital payments by more in districts where pre-interoperability fragmentation is higher: - ∂∆N_D_d / ∂F_d > 0. - Prediction 2 (Interoperability and usage per platform): - When pre-interoperability fragmentation is relatively low, introducing interoperability increases usage on each platform by more in districts where pre-interoperability fragmentation is higher: - ∂∆N_d,a / ∂F_d > 0 and ∂∆N_d,b / ∂F_d > 0. ### 3.4 Discussion of modeling choices - Focus on adoption and number of users (not transactions): - Equilibrium concept based on numbers of users; total transactions for each method proportional to its number of users. - Empirical analysis disaggregates three margins: (i) average value per transaction, (ii) number of transactions per digital payment user, (iii) number of users per capita. Margin (iii) accounts for the majority of observed variation. - Focus on users rather than platforms: - Model abstracts from platform provider objectives and actions; interoperability is exogenously imposed or not. - Unification via interoperability delivers same total adoption as pooling on a single private platform in absence of heterogeneous valuations, while preserving user choice. - Provider behavior, rent extraction, and innovation incentives are abstracted from and left for future work. - Within-district payments only: - Cross-district payments analyzed in Appendix F. - Presence of cross-district transactions would attenuate estimated impact of integration, implying empirical estimates likely a lower bound. - Exogenous and discrete preference boundary ˆx_d: - ˆx_d is assumed exogenous and unaffected by interoperability, simplifying equilibrium derivation. - This removes a “platform switcher” margin; incorporating it would not change main empirical predictions if potential switchers are few. *Source: wpiea2025126-print-pdf - 3.2 Equilibrium analysis (canonical URL: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025126-print-pdf.pdf)* ### 4.1 Estimation strategy ### 4.1 Estimation strategy ### Variation and empirical measure of fragmentation - Empirical setting: integration with UPI of a major incumbent fintech firm that previously offered only closed-loop payments; incumbent’s platform had a total transaction value similar to UPI in the month prior to integration, but UPI was growing substantially faster. - Geographic scope: total transaction value on UPI was higher than that on the incumbent’s platform in 97% of districts in the month before integration. - Fragmentation measure Fd (district d, period t−1): - Fd := Total value of transactions per capita on the smaller platform in district d in t−1 / Total value of transactions per capita across both platforms in district d in t−1. - Interpretation: Values closer to 50% imply higher fragmentation across platforms; values closer to zero imply consolidation on one platform. - Empirical distribution: median(Fd) = 7.4%. - For regressions, construct dummy F+d = 1 for districts with Fd above the median. ### Specification (heterogeneous adoption design) - Estimation window: six months before integration to one year after. - Baseline regression (district-month level): - ydt = αd + αst + β(F+d × 1{t≥t0}) + βZ (Zyd × 1{t≥t0}) + edt - αd and αst are district and state-time fixed effects. - 1{t≥t0} indicates post-integration periods. - Zyd controls for pre-integration digital payments usage; baseline Zyd = overall value of transactions per capita across both platforms in t−1 (the denominator in equation (3)). - Standard errors clustered at the district level. - Top 1% of ydt winsorized in the baseline specification. - Event-study specification: - ydt = αd + αst + Στ≠t−1 βτ (F+d × 1{t=τ}) + Στ≠t−1 ατ (Zyd × 1{t=τ}) + vdt - Month-wise dummies estimate dynamics relative to the pre-integration baseline. ### Outcomes and exclusion of P2P - Primary outcome: total peer-to-merchant (P2M) transaction value across both platforms per capita (P2P transactions excluded to focus on within-district payments and to eliminate remittances). ### Identification concerns and mitigations - Anticipation: - Concern: users could have anticipated integration, biasing β. - Mitigations: - Integration decision immediately preceded by an RBI directive mandating wallets make their wallets interoperable through UPI; accurate widespread anticipation deemed implausible. - If anticipation occurred, it would bias β downward, making estimated effects conservative. - Event-study shows no differential pre-trends between high- and low-Fd districts. - Parallel trends and confounders: - Five mitigations: 1. State-time fixed effects compare only within-state variation. 2. Control for differential trends by pre-integration digital payments usage (Zyd). 3. No statistically significant pre-integration trend differences (from event-study). 4. Results robust when comparing to a matched sample of low-Fd districts similar on observables. 5. Results robust when instrumenting Fd with geographic decisions by the incumbent taken more than a year prior to integration (before the demonetization shock). ### Footnote/auxiliary details - Cross-district transactions discussion referenced to Section 3.4 and Appendix F. - Decomposition procedure described in Appendix G. --- ### Key estimation parameters and design choices - Estimation sample: six months before to one year after integration. - Clustering: standard errors clustered at the district level (Bertrand, Duflo, and Mullainathan, 2004). - Winsorization: top 1% of ydt winsorized in baseline. - Matching: Mahalanobis distance matching on log population with replacement; retained 500 districts; discarded 54 low-Fd districts never matched and 19 high-Fd districts with populations outside low-Fd range. - Instrumental variable: define H d as negative distance in kilometers from district centroid to nearest “hub” centroid; identify eight hub districts with at least 1000 merchants adopted by September 2016; distributional facts about merchant counts pre-2016: 92% of districts had fewer than 100 merchants signed up, 85% had fewer than 10, and 42% had zero. *Source: wpiea2025126-print-pdf - 4.1 Estimation strategy* ### 5.1 Aggregate national impact ### 5.1 Aggregate national impact ### Model objective and identification challenge - Object of interest: ∆I_ND, the total national change in usage of digital payments that results from integrating fragmented networks. - Identification challenge: the “missing intercept” problem — unknown absolute impact of network unification on districts that were less fragmented ex ante. - Solution approach: impose additional structure from the model (extended model in Appendix D) to derive absolute impacts for each district and aggregate to national level. - Key theoretical result (Proposition 5): ∆I_ND = Σ_d (∆N^D_d − ∆N^D_d0), where ∆N^D_d0 is the post-interoperability change in a district whose digital payments users are already fully unified prior to interoperability (equation (7)). - Model implication used: connecting the two networks has no impact on a district that is already fully unified on one platform ex ante, so ∆N^D_d0 = ω and observed changes in those districts can net out unobserved shocks in other districts. ### Empirical strategy to recover absolute impacts - Proxy for “fully unified ex ante” districts d0: first decile of the F_d distribution (these districts have a median F_d of 1%). - Estimation specification (relative-decile baseline, equation (8)): - y_dt = α_d + α_st + Σ_{n=2}^{10} β_n (F^n_d × 1{t≥t0}) + β_Z (Z^y_d × 1{t≥t0}) + e_dt - F^n_d is a dummy for district d in the nth decile of F_d. - Omitting interaction for F^1_d makes β_2,...,β_10 estimate ∆N^D_d − ∆N^D_d0 within decile n. - Outcome y_dt: total P2M transaction value per capita (as in Section 4.1). - Interpretation: each estimated coefficient β̂_n is the monthly average increase in total P2M transaction value per capita after interoperability, for districts in decile n, relative to first-decile districts, controlling for district and state-time fixed effects and differential trends by total pre-integration transaction value. - Robustness note: Appendix Figure A.9 plots coefficients; observed pattern matches model prediction (limited impact in initially unified districts, increasing impact with higher initial fragmentation). - Sensitivity to choice of d0: repeating aggregation using more or fewer quantiles yields estimates remaining between 40% and 65% (see footnote 38 and Appendix Figure A.10). ### Aggregation to national impact - Empirical analogue of equation (7) for population-weighted national average (equation (9)): - ∆I_y = Σ_d Population_d / Σ_d Population_d × Σ_{n=2}^{10} β̂_n × F^n_d - Procedure: calculate β̂_n × F^n_d for each district and decile, aggregate using district population weights, then re-normalize by national population. ### Key numerical findings and interpretation - Estimated population-weighted national average monthly increase in total P2M transaction value per capita attributable to interoperability: - ∆I_y = 10.2 Rupees per capita per month during the first year after integration. - Counterfactual comparison and relative increase: - Integration raised average national P2M transaction value per capita per month by 59% relative to the model-implied value that would have occurred in the absence of the integration event (equation (10)). - Interpretation and scope: - Aggregate result indicates a substantial impact of integration on overall usage of the payments system. - Result accounts for the total reduced-form effect of network integration, including both transactions conducted directly between the two platforms and any additional activity induced by expansion of the accessible network. - Given UPI integrated additional pre-existing networks (none as large as the incumbent platform studied), findings suggest interoperability played a significant role in the system’s take-off. - Caveats: - Aggregation inherits the model assumptions and trade-offs discussed in Section 3.4; estimates may under-state the true impact (e.g., some users in “fully unified ex ante” districts could still gain from cross-district transactions, which would mean netting out some increases that actually result from interoperability). ### Downstream lending spillovers (summary of Section 5.2) - Research question: Do higher payments usage after integration lead to more borrowing? - Specification at household-survey wave level (equation (11)): - y_ht = α_h + α_st + β (F^+_d × 1{t≥t0}) + β_Z (Z^y_d × 1{t≥t0}) + e_ht - Outcome: probability of borrowing from non-banks (non-bank fintech firms more likely to use payments-generated information). - Controls: household fixed effects α_h, state-time fixed effects α_st, controls for differential trends by overall digital payments usage ex ante; standard errors clustered at district level. - Empirical findings: - After integration, in districts with above-median fragmentation, the probability of borrowing increases by 1.1 percentage points relative to districts with below-median fragmentation. - These increases are larger for households more likely to benefit from digital payments activity (entrepreneurs or hawkers). - Similar results found for households with no previous borrowing activity (Appendix Table A.2). - Interpretation: - Evidence consistent with interoperability-induced increases in retail digital payments reducing frictions in credit markets and enabling more households to borrow. ### Long-term supply-side considerations (summary of Section 5.3) - Long-run ambiguity: interoperability can create coordination effects (investments complementary across firms) and free-rider effects (under-investment because gains accrue to others). - Empirical literature notes potential for interoperability to weaken providers’ incentives to expand infrastructure coverage in some settings; other evidence shows data-sharing can induce net complementarities in investment. - In the UPI context: - Evidence of endogenous provider responses to re-create proprietary within-app network effects despite interoperable rails: - Proprietary QR codes initially used (later prohibited by RBI). - Branding of interoperable QR codes that suggested app-specific readability. - Referral schemes incentivizing within-app transactions. - Observed patterns: - In roughly half of UPI transactions the payer and payee continue to use the same app; ratio of cross-app to within-app transactions remained broadly stable in recent years (Figure 4). - Increasing regional concentration over time in user bases of the two largest apps (Appendix Figure A.11). - Policy implication: - Maintaining a truly interoperable system may require continuing regulatory and policy efforts to counter provider strategies that re-introduce lock-in and proprietary network effects. - A fuller assessment of long-term supply-side impacts and welfare implications is left for future work as more data on providers’ investments and innovations become available. *Source: wpiea2025126-print-pdf - 5.1 Aggregate national impact* ### References ### References ### Major themes and literature areas - Digital payments adoption and interoperability (UPI, mobile money, fast payments). - Network effects, platform competition, and two-sided markets. - Monetary economics: currency competition, digital currencies, CBDC, and international currency dynamics. - Empirical methods: differences-in-differences, natural experiments, high-resolution geographic analysis, and instrumental variables. - FinTech lending, shadow banking, and credit implications of payment data sharing. - Policy and regulatory perspectives on payment systems, competition, and cross-border payments. ### Representative works and details (selected entries from the reference list) - ACI WORLDWIDE (2023): “Prime Time for Real-Time Report 2023,” Industry report. - ADRIAN, T., F. GRINBERG, T. MANCINI-GRIFFOLI, R. M. TOWNSEND, AND N. ZHANG (2022): “A Multi-Currency Exchange and Contracting Platform,” Working Paper 217. - AGARWAL, S., P. GHOSH, J. LI, AND T. RUAN (2024): “Digital Payments and Consumption: Evidence from the 2016 Demonetization in India,” TheReviewofFinancialStudies. - ALONSO, C., T. BHOJWANI, E. HANEDAR, D. PRIHARDINI, G. UNA, AND K. ZHABSKA (2023): “Stacking Up the Benefits: Lessons from India’s Digital Journey,” Tech. rep., International Monetary Fund, Washington, D.C. - ALVAREZ, F. E., D. ARGENTE, F. LIPPI, E. MÉNDEZ, AND D. V. PATTEN (2023): “Strategic Complementarities in a Dynamic Model of Technology Adoption: P2P Digital Payments,” NBERWorkingPapers, 31280. - ARGENTE, D., P. GONZALEZ ALVAREZ, E. MÉNDEZ, AND D. VAN PATTEN (2025): “Drivers of Digital Payment Adoption: Lessons from Brazil, Costa Rica, and Mexico,” Working Paper 34280, National Bureau of Economic Research. - BIANCHI, M., M. BOUVARD, R. GOMES, A. RHODES, AND V. SHREETI (2023): “Mobile Payments and Interoperability: Insights From the Academic Literature,” InformationEconomicsandPolicy, 65, 101068. - BRUNNERMEIER, M. K., N. LIMODIO, AND L. SPADAVECCHIA (2023): “Mobile Money, Interoperability, and Financial Inclusion,” SSRN Scholarly Paper 4574641. - FINANCIAL STABILITY BOARD (2025): “G20 Roadmap for Cross-Border Payments: Consolidated Progress Report for 2025,” Tech. rep. - GOVERNMENT OF INDIA (2026): “Socio-Economic Impact Analysis of Incentive Scheme for Promotion of RuPay Debit Card and Low-Value BHIM-UPI Transactions (P2M),” Report, Department of Financial Services, Ministry of Finance, Government of India. - RESERVE BANK OF INDIA (2016): “Payment and Settlement Systems in India: Vision-2018,” Vision Document. - WOLF, C. K. (2023): “The Missing Intercept: A Demand Equivalence Approach,” AmericanEconomicReview, 113, 2232–2269. - WOOLDRIDGE, J. M. (2015): Introductoryeconometrics:Amodernapproach, Boston, MA: Cengage Learning, 6th ed. ### Figures and tables: highlights and key numeric findings - Figure 1: Cumulative apps and banks participating in UPI (time series plotted from 2016m1 to 2024m1). - Figure 2: Electronic retail payments in India — value (percent of GDP) and volume (transactions per capita) from 2012 to 2024; data sources include RBI, NPCI, Haver Analytics, WDI. - Figure 3: Volume of fast payment transactions (millions) across countries (2012–2022); log-scale categories include 10, 100, 1000, 10000, 80000. - Figure 4: Share of cross-app transactions on UPI (%) plotted from 2017m1 to 2024q1; methodology accounts for unobserved “Other app” to “Other app” cell with Appendix I procedure. - Figure 5: Indexed comparison of Closed-loop incumbent, UPI, and ATM withdrawals around demonetization (2016m10 = 100); state-level median with 25-75th and 10-90th percentiles shown. - Figure 11: Response dynamics of total digital payments adoption to platform integration — Panel (a) Total P2M transaction value per person (Rupees per capita) and Panel (b) value relative to cash withdrawals — plotted for months since integration (range -6 to 12); 95% confidence intervals shown. - Figure 12: Cross-platform transactions after platform integration — Panel (a) response dynamics of cross-platform P2M transaction value (Rupees per capita) and Panel (b) cross-sectional variation one year after integration vs. ex-ante fragmentation. - Table 1: Response of digital payments adoption to platform integration (district-level regressions; sample spans six months before integration to one year after). Key coefficient estimates (F + d × 1{t≥t0}): - Column (1) Total/pop: 8.342 ∗∗∗ (t-stat 4.67) - Column (2) Incumbent/pop: 4.925 ∗∗∗ (t-stat 6.03) - Column (3) Other/pop: 2.971 ∗∗∗ (t-stat 3.70) - Column (4) (Inc↔Oth)/pop: 0.208 ∗∗∗ (t-stat 4.13) - N: Columns (1)–(3) N = 10,868; Column (4) N = 7,436. - Mean y dt (F + d = 1,t = t −1 ): 9.018, 7.089, 1.928 (Columns 1–3). - Mean y dt (F + d = 0,t≥ t 0 ): 6.687, 1.450, 5.294, 0.189 (Columns 1–4 as applicable). - Table 2: Matching on log population size; F + d × 1{t≥t0} coefficients: - Column (1): 7.002 ∗∗∗ (t-stat 4.82) - Column (2): 4.524 ∗∗∗ (t-stat 6.40) - Column (3): 2.364 ∗∗∗ (t-stat 3.16) - Column (4): 0.172 ∗∗∗ (t-stat 3.41) - N: Columns (1)–(3) N = 9,500; Column (4) N = 6,500. - Table 3: Instrumenting with proximity to incumbent hub districts (2SLS); F + d × 1{t≥t0} coefficients: - Column (1): 17.76 ∗∗∗ (t-stat 2.74); K-P F -Stat 20.44 - Column (2): 10.01 ∗∗∗ (t-stat 2.97); K-P F -Stat 19.88 - Column (3): 6.311 ∗∗ (t-stat 1.98); K-P F -Stat 21.43 - Column (4): 0.642 ∗∗ (t-stat 2.37); K-P F -Stat 3.15 - N: Columns (1)–(3) N = 10,621; Column (4) N = 7,267. - Table 4: Household-level NBFC borrowing response (household panel; waves every four months); F + d × 1{t≥t0}: - Column (1) All: 0.0113 ∗∗ (t-stat 2.17), N = 898,412 - Column (2) Entrepreneurs: 0.0192 ∗∗ (t-stat 2.54), N = 54,161 - Column (3) Hawkers: 0.0136 ∗∗∗ (t-stat 3.00), N = 22,387 - Appendix figures and tables provide robustness, matching, instrument validity, decomposition, and additional dynamics (e.g., Figure A.1–A.11, Table A.1–A.2). ### Appendices and supplementary analyses (structure and focus) - Online Appendix contents: - A Additional figures and tables - B Cross-checking our location proxy - C Proofs - D Extended model allowing for external shocks - E Implications of potential multihoming - F Implications of cross-district transactions - G Decomposition of the change in value per capita - H Additional robustness checks - I Estimating cross-app “Other-Other” transactions - Key methodological points in appendices: - Location proxy cross-checks: correlation coefficient 0.84 between NPCI-mapped PhonePe transactions and PhonePe Pulse (Figure B.1a). - Pincode-level exposure measure Exposure_p defined in equation (B.1). - Theoretical model extensions include an unanticipated shock ω with 0 < ω < 1 − 1/(γ−κ) (Appendix D). - Estimation strategy for unobserved “Other-Other” cross-app transactions uses a φ ratio approach (Appendix I, equations I.1–I.6). *Source: wpiea2025126-print-pdf - References (Working Paper PDF).*/ --- _Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025126-print-pdf.pdf_