Avanti — Collections Forecast

Peak collections

₹133.2 Cr

Jul 2025

Last actual month

₹92.0 Cr

Apr 2026

Forecast — May 2026

₹88.9 Cr

Next month

6-month forecast total

₹496.7 Cr

May – Oct 2026

Peak → Oct 2026

−43%

Jul 25 → Oct 26

Collections trend — actual & forecast

All 26 states · click any tab or state card below

Actual collections TimesFM forecast 90% confidence band

Select a state above to see analysis.

All 26 states — at a glance

6-month forecast total · sorted by volume · click any card to drill in

Decision signals

What this data is telling you

▼

UP & West Bengal — structural breaks

UP forecast to fall to ₹6 Cr by Oct 2026 — a 62% drop from peak. WB dropped ₹4 Cr in a single month (Oct 2025) and has not recovered. Both have wide confidence bands. Ground-level review needed immediately.

Rajasthan, Gujarat & Haryana — near-zero horizon

Rajasthan forecast at ₹0.68 Cr, Gujarat ₹0.82 Cr, Haryana ₹0.29 Cr by Oct 2026. If active borrowers remain, this is a collections failure in the western and northern clusters requiring immediate field action.

The Oct 2025 national step-down

Collections fell ~₹20 Cr nationally in a single month across multiple states simultaneously. A portfolio-level event absorbed as a structural break in every affected state. Understanding the cause is key to validating the forecast direction.

Tamil Nadu & MP — spikes were not trends

TN spiked to ₹16.5 Cr in Jul 2025. MP to ₹19.4 Cr in Sep 2025. Both correctly treated as noise. If campaign-driven, the question is whether they can be replicated to shift the forecast trajectory.

↔

Assam, MP & Jharkhand — hold the line

Stable or flat trajectories with tighter confidence bands. Assam most consistent for 12 months. Deploying collection resources here has predictable return on effort and lower risk of further deterioration.

This forecast assumes nothing unusual happens

TimesFM only sees historical numbers — not monsoon stress, campaigns, policy changes, or restructuring. Use the point forecast as your base case. Lower confidence bound = stress scenario; upper = upside if field performance improves.

How this was built

Model, data, and appropriate use

Model

Google TimesFM 2.5 — a 200M parameter decoder-only transformer pre-trained on 400+ billion real-world time points. Runs natively in BigQuery via AI.FORECAST(). No training required. Zero-shot: predicts your data without ever having seen MFI collections before.

Data

Monthly amounts aggregated from paymentLedger in Latest_Avanti_Loan. Reversed payments excluded. Current partial month excluded. 100 months of history (Feb 2018 – Apr 2026). All 26 states forecasted as independent time series. ~320 MB scanned per run.

Confidence band

The shaded region is a 90% prediction interval. Bands widen further out as uncertainty compounds. Wide bands (UP, TN) mean the model is genuinely uncertain. Tight bands (Assam, Mizoram) mean it is confident. Treat lower bound as stress scenario, upper as optimistic.

Limitations

TimesFM does not know about monsoon patterns, waiver announcements, new campaigns, NPA restructuring, or one-off drives. Spikes are treated as noise. Kerala has insufficient recent data — exclude from planning. For scenario planning, apply judgment multipliers on top of the confidence band bounds.