Collection Forecasting for NBFCs: Bucket-Level Methods
Last reviewed: November 2026
Collection forecasting for NBFCs is the practice of predicting how much money a lender will collect in a coming week or month, split by delinquency bucket, branch and payment channel. Most non-banking financial companies (NBFCs) still roll forward last month's efficiency, so treasury and field teams plan on a number that misses badly in festival and harvest months.
This guide is for chief financial officers (CFOs) and heads of collections at NBFCs and fintech lenders. It covers the forecasting methods, the inputs, a worked example, liquidity and model governance, and how Digiqt's Custom AI Development team can build an artificial intelligence (AI) forecasting model for your book.
Key Takeaways
- Forecast by bucket first. Each days past due (DPD) bucket collects at a very different rate.
- Split by branch and channel next, because mandate inflows and field cash behave differently.
- Link the forecast to liquidity. Repayment inflows feed the structural liquidity statement.
- Forecasting can be systematic. A validated model can refresh weekly and leave analysts the overrides.
What Is Collection Forecasting for NBFCs?
Collection forecasting for NBFCs is a book-level estimate of future inflows, used by finance, treasury and collections. It answers three questions: how much cash arrives, when, and through which channel.
It is different from measuring past results, which our guide to collection efficiency covers, and from scoring single borrowers, which self-cure prediction in collections handles. A forecast adds those borrower-level signals up into one planning number.
The CFO uses it for borrowing and liquidity buffers. The head of collections uses it to size field and calling teams, alongside collections workforce planning.
What Does a Last-Month Forecast Cost an NBFC?
A last-month forecast costs an NBFC idle liquidity, because treasury must hold a larger buffer against forecast error. The illustrative table below assumes ₹300 crore of monthly instalment demand and a 3% negative carry on surplus cash.
| Measure | Last-Month Run Rate | Bucket-Level Forecast |
|---|---|---|
| Typical forecast error | ±4% | ±1.5% |
| Error in rupees | ₹12 crore | ₹4.5 crore |
| Buffer held against error | ₹12 crore | ₹4.5 crore |
| Annual negative carry at 3% | ₹36 lakh | ₹13.5 lakh |
| Annual saving | – | ₹22.5 lakh |
The larger cost is operational. Field teams sized for an average month are short in harvest delays and idle in salary weeks.
How Does a Bucket-Level Collection Forecast Work?
A bucket-level collection forecast works by applying an expected collection rate to the demand in each DPD bucket and adding the results. The rates come from recent roll rates, the same month last year and current payment signals.
Under the Reserve Bank of India (RBI) NBFC Income Recognition, Asset Classification and Provisioning Directions, 2025 (RBI/DOR/2025-26/356, 28 November 2025), an NBFC in the middle layer treats an asset as non-performing (NPA) once interest is overdue for more than 90 days. That is why most forecasts stop the flow at the 90 DPD line and model NPA recoveries separately.
Which Inputs Drive a Collection Forecast?
Six inputs drive a collection forecast: demand schedule, mandate status, roll rates, seasonality, field capacity and local events.
| Input | Why It Matters | Source |
|---|---|---|
| Demand schedule | Sets the maximum collectable | Loan management system (LMS) |
| Mandate status and bounce history | Predicts auto-debit success | National Automated Clearing House (NACH) and Unified Payments Interface (UPI) AutoPay files |
| Bucket roll rates | Predicts cures and slippages | Monthly DPD movement |
| Seasonality | Harvest, festival and salary-date effects | Same month in prior years |
| Field capacity | Caps cash collection in late buckets | Collections roster |
| Local events | Floods, elections, strikes | Branch inputs |
Bounce history is often the strongest early signal, which is why NACH bounce rate analysis belongs in the data pipeline.
Should NBFCs Use Roll Rates or Machine Learning?
NBFCs should use roll rates as the baseline and add machine learning (ML) where the book is large and varied enough. The choice depends on data depth and how much of the book behaves seasonally.
| Approach | Best For | Weakness |
|---|---|---|
| Roll-rate (flow) model | Stable, single-product books | Misses seasonality and mix shifts |
| Vintage-based curves | Fast-growing books | Needs several cohorts of history |
| Machine learning | Multi-product, multi-state books | Needs validation and explainability |
| Hybrid | Most mid-sized NBFCs | More moving parts to govern |
Cohort curves from vintage analysis of the loan portfolio are a practical input for growing books.
How Should NPA Recoveries Be Forecast?
NPA recoveries should be forecast separately from the bucket flow, because they depend on legal stages, settlements and repossessions rather than monthly payment habits. A practical method lists expected recoveries case by case for large accounts and applies a historical recovery rate to the rest.
Treatment choices, set out in our guide to bucket-wise collection strategy, change these rates. Rerun the forecast whenever the strategy for a bucket changes.
How Do You Split Collection Forecasting by Branch and Channel?
Collection forecasting by branch and channel splits the bucket total into five channel streams, each with its own driver.
| Channel | Main Driver | Forecasting Note |
|---|---|---|
| NACH auto-debit | Presentation dates, bank bounce rates | Most predictable stream |
| UPI AutoPay | Mandate health, balance on debit day | Compare with NACH behaviour |
| Payment links | Reminder timing | Peaks after reminders |
| Field cash | Agent capacity, route plans | Constrained by staff |
| Branch walk-in | Market days, harvest cash | Strongly seasonal |
Channel mix shifts as lenders move borrowers between mandates, as our comparison of eNACH vs UPI AutoPay explains. Branch forecasts then roll up to region and state for the CFO.
Is your treasury still planning next month on last month's efficiency?
See How Digiqt Builds Custom AI Models
Learn how a discovery sprint scopes a forecasting model on your own data, or speak with our AI engineering team.
Worked Example: One Month, Four Buckets
In this illustrative example, an NBFC with ₹300 crore of demand forecasts ₹259.04 crore of collections for a normal month and ₹254.04 crore for a festival month.
| Bucket | Demand | Expected Rate | Forecast |
|---|---|---|---|
| Current (0 DPD) | ₹250 crore | 95% | ₹237.50 crore |
| 1–30 DPD | ₹30 crore | 55% | ₹16.50 crore |
| 31–60 DPD | ₹12 crore | 30% | ₹3.60 crore |
| 61–90 DPD | ₹8 crore | 18% | ₹1.44 crore |
| Total | ₹300 crore | – | ₹259.04 crore |
History shows the current-bucket rate falls two points in festival months, to 93%, cutting that line to ₹232.50 crore. The ₹5 crore gap tells treasury to plan a smaller inflow and tells collections to shift callers into early buckets.
How Does Custom AI Development Build a Collection Forecasting Model?
Digiqt's Custom AI Development team builds a collection forecasting model by scoping the problem in a discovery sprint, testing feasibility on your data, then selecting, validating and deploying the model. A dedicated development team handles the build and continuous improvement.
Custom AI Development is Digiqt's service for designing and implementing AI solutions, including custom machine learning modelling, explainable AI, and analytics and reporting.
| Task | Manual Today | With a Custom Forecasting Model |
|---|---|---|
| Build forecast | Spreadsheet run rate | Custom ML model chosen after dataset assessment |
| Seasonality | Analyst memory | Learned from prior years' data |
| Branch and channel split | Pro-rata allocation | Forecast at branch and channel level |
| Accuracy check | Rarely done | Model validation and performance analysis |
| Explain to committee | Hard to defend | Explainable AI shows which drivers moved |
| Refresh | Monthly, manual | Deployed with continuous improvement |
Your credit, treasury and collections leaders still own the assumptions, overrides and final approval.
How Does a Collection Forecast Feed Liquidity Planning?
A collection forecast feeds liquidity planning by estimating the loan repayment inflows that sit in the structural liquidity statement (SLS). According to the RBI NBFC Asset Liability Management Directions, 2025 (RBI/DoR/2025-26/355, 28 November 2025), inflows and outflows are placed by expected timing.
The same Directions cap net cumulative negative mismatches at 10% in the 1–7 day bucket, 10% in the 8–14 day bucket and 20% in the 15–30 day bucket, as a share of cumulative outflows.
| Forecast Horizon | Liquidity Use |
|---|---|
| Next 7 days | SLS 1–7 day bucket, daily cash plan |
| 8–30 days | SLS 8–14 and 15–30 day buckets, borrowing calls |
| 1–3 months | Drawdown timing on bank lines |
| 3–12 months | Business plan and fundraising |
Forecasts also help time draws on sanctioned lines, which our guide to NBFC borrowing drawdown tracking covers. Digiqt's liquidity forecasting AI agent page shows the wider treasury view.
How Should NBFCs Govern a Collection Forecasting Model?
NBFCs should govern a collection forecasting model like any other decision model: inventory it, validate it, back-test it and log overrides. RBI's draft Guidance on Regulatory Principles for Model Risk Management, released on 24 June 2026, proposes this for all models used by regulated entities, including NBFCs.
The draft proposes a board-approved model risk framework, a model inventory and independent validation before and after deployment. It was still a draft at the time of review, so treat it as direction of travel.
- Back-test monthly by bucket, branch and channel.
- Set error tolerances by horizon, tighter for next-week NACH than for late buckets.
- Log every override with the reason and who approved it.
- Re-validate after shocks, such as a moratorium or a regional flood.
How Accurate Should a Collection Forecast Be?
A collection forecast should be judged against tolerances set by horizon and segment, not one company-wide figure. Next-week NACH inflows are close to known, while three-month late-bucket recoveries carry far more uncertainty.
Measure error as the absolute gap between forecast and actual, divided by actual, for each bucket, branch and channel. Any segment that misses tolerance two months running needs a review of its inputs, not just another manual override.
Want a collection forecast your board and treasury both trust?
Talk to Digiqt About a Collection Forecasting Model
Learn how the discovery, validation and deployment steps work, or visit Digiqt to schedule a scoping call.
Sources
- RBI: NBFC – Asset Liability Management Directions, 2025
- RBI: NBFC – Income Recognition, Asset Classification and Provisioning Directions, 2025
- TaxGuru: RBI Issues Draft Guidance on Regulatory Principles for Model Risk Management
- Analytics India Magazine: What RBI's AI Model Risk Draft Means for Lenders
- TaxGuru: Summary of NBFC IRACP Directions, 2025



