Point-in-time feature serving without temporal data leakage

FEATSRV · Actuarial Feature Store Research Platform

Download paper

Zi Humana Research presents FEATSRV — a leakage-free point-in-time feature store for insurance. Feast, Redis, PySpark, and FastAPI deliver SCD Type 2 as-of joins, online serving, and the Nimbus actuarial benchmark for underwriting, pricing, and fraud models.

10.5281/zenodo.21798383Vipul JainFeatLock 1.0
Online lookup latency
<10 ms
Redis policy & claim feature fetch
PIT join engines
2
Feast/Dask · PySpark SQL equivalence
REST surface
8
Health, online, offline, PIT, audit, FI
Nimbus packs
7
SCD Type 2 traps for leakage evaluation

Our work

Point-in-Time Feature Serving in Actuarial Data Science: Foundations, Leakage Dynamics, and FEATSRV

Research Paper
Vipul Jain · Zi Humana · AE Quantum Research Division

This report documents FEATSRV: bitemporal theory and PIT as-of joins for SCD Type 2 actuarial entities, dual Feast/Dask and PySpark engines with byte-identical regression checks, eight FastAPI routes for online Redis serving and leakage audits, and the Nimbus synthetic benchmark with intentional temporal traps for underwriting, pricing, and fraud workflows.

Point-in-TimeFeature StoreActuarial ML

Administrative console overview — uptime and request volume

Figure 1
Health, 24h traffic, and response-time distribution
OpsKPIs

Data lineage from Parquet sources through PIT store to models

Architecture
Feast registry metadata · end-to-end audit path
FeastLineage
Open access archive on ZenodoRead more research on our hub

Results & figures

Table 1. FEATSRV REST surface
EndpointMethodRole
/healthGETUptime & Redis / Feast dependency checks
/online/policy/{id}GETReal-time policy features from Redis
/online/claim/{id}GETReal-time claim features from Redis
/pit-trainingPOSTPoint-in-time join on uploaded target CSV
/leakage-auditPOSTScan datasets for temporal leakage traps
Table 2. Five-layer system architecture
LayerStackFunction
DataParquet · PySpark · SCD Type 2Temporal policies, claims, and addresses
Feature storeFeast · SQLite · RedisRegistry, offline joins, online materialization
ServingFastAPI · Uvicorn · DaskREST for online, PIT, audit, and importance
ClientCLI · HTML consoleOperator workflows and dashboard panels
FEATSRV console overview panel with KPIs and request volume
Figure 1. Overview panel — system uptime, 24-hour request volume, and response-time distribution from the FEATSRV administrative console.
FEATSRV data lineage from temporal sources to model consumers
Figure 2. Data lineage — Parquet temporal sources through the PIT store and Redis to actuarial model consumers.
FEATSRV risk and leakage audit panel
Figure 3. Risk & leakage audit — portfolio risk index, entity distributions, and identified temporal leakage traps from controlled Nimbus fixtures.
Feature importance panel across actuarial models
Figure 4. Feature importance — attribution scores across underwriting, pricing, and fraud modeling tasks.
Point-in-time training and batch set panel
Figure 5. Training & batch sets — PIT training previews and offline join logs before model fit.
Online Redis feature lookups for policy and claim entities
Figure 6. Online lookups — live Redis key-value payloads for policy and claim entity IDs.
Latency by FEATSRV service microbenchmark chart
Figure 7. Microbenchmark latency by service — online Redis lookups versus Dask and PySpark offline PIT join workloads.

Open access · FEATSRV Research Initiative

Vipul Jain · Zi Humana · doi:10.5281/zenodo.21798383 · Point-in-Time Feature Serving · Borel Sigma Inc. venture

Download full report

Tools + Code

Research Paper

FEATSRV engineering report — point-in-time joins, leakage dynamics, Nimbus benchmark, and dual Dask/PySpark engines.

View project →

FeatLock 1.0

Feature serving dashboard — lineage, leakage audit, feature importance, PIT training, and online Redis lookups.

View project →

spark_pit_join.py

PySpark SQL point-in-time join engine with ROW_NUMBER windowing for portfolio-scale training sets.

View project →

Nimbus Benchmark

Synthetic actuarial SCD Type 2 dataset with intentional leakage traps across seven service packs.

View project →