PACE Racing Analytics
Thoroughbred racing data pipeline · 2000 to present
AI powered sports data engineering · Predictive analytics
AI powered thoroughbred racing analytics platform: reconciling three independent providers that share no identifiers, delivering AI driven sectional analysis from 25 Hz GPS telemetry across a 25 year archive, with predictive modelling and anomaly detection trained on decades of racing data.
25 yrs
Archive depth: 2000 to present
3
Independent data providers reconciled
~25 Hz
GPS telemetry sampling per runner
The Challenge
Thoroughbred racing analytics is a multi source reconciliation problem before it is an analytics problem. Three independent providers describe the same race: none of them agree on identifiers, none arrive on the same clock, and one (tracking) produces four orders of magnitude more data than the others. On top of that sits a 25 year archive where data fidelity degrades sharply backwards in time. The pipeline must ingest, reconcile, and derive performance analytics across all racing jurisdictions from 2000 to present.
Data Sources
PA (Press Association)
Racecards, declarations, runners, riders, weights, going, non runners, results, form, historical archive
Scheduled feed drops plus intraday deltas: authoritative for entity master data
Course Track
Positional tracking data per runner
Per race, post race or near live delivery
GMAX / Race IQ (Total Performance Data)
Sectional timing, in running positional data, GPS telemetry from saddle cloth trackers
Live streaming during running plus post race settled files at ~25 Hz
What We Built
An AI powered, event driven pipeline that reconciles three independent feeds into a canonical racing datastore, derives telemetry grade analytics from raw GPS using machine learning models, and serves the result through a Fastify API + Vue.js front end purpose built for large numeric payloads: with AI driven anomaly detection, predictive performance modelling, and automated data quality assurance across the entire archive.
AI powered entity resolution
Three providers, zero shared identifiers. AI driven canonical registry maps provider native IDs via deterministic matching on strong keys (course + date + off time + distance) and ML based fuzzy matching with confidence scoring on horse and jockey names (handling accents, suffixes, country codes, and ownership transfers. Unresolved matches quarantine for review) never silently dropped.
RabbitMQ event backbone
Ingest, transform, and load stages fully decoupled. Topic exchange with structured routing keys, idempotent consumers with natural key upserts, dead letter exchange with bounded retries, and isolated lanes so archive backfill never starves live raceday.
PostgreSQL tiered storage
Reference/form tier (modest volume, deep joins, 25 year history), sectional tier (split times and pace figures), and raw tracking tier (hundreds of millions of rows/year). Date range partitioning, BRIN indexes, materialized views for expensive aggregates.
AI sectional analysis engine
Raw GPS projected into course local planar frame, resolved against modelled track centreline spline. AI enhanced Kalman smoothing recovers velocity without amplifying noise. ML models derive max sustained speed, stride metrics, ground loss, and per furlong sectionals with predictive accuracy beyond manual analysis.
AI normalisation & predictive modelling
Raw splits normalised against AI generated par times by course, distance, going, and class. Machine learning models produce finishing speed percentage, energy distribution, pace ratings, and predictive performance forecasts: comparable across meetings, surfaces, and seasons.
Video alignment
Race replays ingested and aligned to the sectional timeline via race start timestamp offset, so a sectional gate crossing maps to a frame.
AI anomaly detection & forecasting
ML models trained on the 25 year archive detect data anomalies (GPS dropout, sensor drift, stewards' revisions) in real time and flag inconsistencies before they reach consumers. Predictive models forecast race performance, surface distance suitability, and identify form patterns invisible to manual analysis.
Outcome
A live AI powered platform at pace.mysportsanalysis.com serving form analysts, racing professionals, and data consumers globally. AI models trained on the full 25 year archive deliver predictive performance forecasting, automated video frame analysis, and real time anomaly detection , capabilities that were previously impossible with manual analysis alone. The pipeline handles live raceday spikes, full archive reprocessing, and everything between: with the same code path for both.
Quick facts
Platform
PACE: MySportsAnalysis
Live at
pace.mysportsanalysis.com
Domain
Thoroughbred racing analytics
Coverage
All jurisdictions: global
Archive
2000 to present (mixed fidelity)
Category
Sports data engineering · AI analytics
Stack
Node.js · Fastify · Vue.js · PostgreSQL · RabbitMQ · Redis · AWS · AI/ML
Build the next one
Same pipeline architecture, same engineering standard: for your sports analytics, IoT, or real time data platform.
Schedule a callSectional analysis engine
From raw telemetry to figures a form analyst can use.
Raw GPS is projected into a course local planar frame, resolved against a modelled track centreline spline, and smoothed with signal processing methods that preserve genuine surges while rejecting sensor noise. The result:
01
Max sustained speed
Peak of a rolling 1 second window on the smoothed velocity curve: not a single sample max. Single sample peaks report GPS dropout artifacts as world records.
02
Stride frequency & length
Extracted from periodic oscillation in velocity/vertical displacement signal via spectral analysis in the galloping band (~2.0 to 2.6 Hz). Stride length = v / f.
03
Ground loss
Total distance travelled vs. nominal race distance yields metres lost racing wide on bends: separates the horse that ran fastest from the one that ran shortest.
04
Sectional splits
s(t) curve inverted to t(s) and interpolated at fixed distance to finish gates (per furlong / 200m): consistent regardless of beacon placement.
05
Pace ratings
Normalised against par times by course, distance, going, and class. Comparable across meetings, surfaces, and seasons.
06
Energy distribution
How each runner deployed effort across the race: front loaded, even, or late. The signal that tells you whether a closing performance was sustainable or a one off.
AI powered analysis
AI that sees what manual analysis cannot.
Machine learning models trained on 25 years of racing data and telemetry deliver capabilities that transform raw numbers into actionable intelligence: from automated video frame analysis to predictive performance forecasting and AI driven trip reconstruction.
AI · 01
AI video frame analysis
Machine learning models analyse race replay footage frame by frame: detecting running position, stride patterns, jockey posture, and tactical decisions in real time. Automated timestamping links visual events to telemetry data for cross referenced analysis.
AI · 02
Predictive performance modelling
AI models trained on 25 years of form, sectional, and telemetry data forecast race outcomes, identify horses whose form cycle suggests imminent improvement, and surface distance suitability signals invisible to manual handicapping.
AI · 03
Anomaly detection & data quality
ML models detect GPS dropout, sensor drift, and stewards' revisions in real time: flagging inconsistencies before they reach consumers. Automated quality scoring on every data point across the entire pipeline.
AI · 04
Running style classification
AI classifies each runner's tactical profile (front runner, stalker, closer, off the pace) from telemetry patterns, tracking how style shifts with distance, ground, and class. Enables race shape modelling and pace scenario simulation.
AI · 05
Trainer & jockey pattern analysis
Machine learning surfaces statistical patterns across connections: which trainers improve horses on first time equipment changes, which jockeys extract the most from closers, where stable form trends precede individual breakthroughs.
AI · 06
AI powered trip analysis
Combining ground loss, sectional data, and video frame analysis, AI reconstructs each horse's actual race experience (trouble in running, wide passages, blocked runs) and adjusts raw performance figures to reveal true ability versus trip luck.
Archive design
Mixed fidelity by design,
not by compromise.
Form, results, and reference data extend to 2000. Tracking and sectional data exist only from provider deployment at each course: roughly mid 2010s onward, uneven by jurisdiction. Modelling that honestly was a core design constraint.
Design consequences
- ✓Data availability is a first class, queryable attribute: not a null
- ✓Analytics degrade gracefully by tier: a 2004 race still returns form based figures
- ✓Backfill is checkpointed, resumable, and replays the identical consumer code as live ingest
Infrastructure
AWS, serverless where it matters.
Racing traffic is violently spiky: raceday concurrency bears no resemblance to 03:00 on a Tuesday. Lambda absorbs the spike and costs nothing between meetings. Redis carries the hot path caching so raceday reads never touch Postgres.
API Gateway
Token validation, throttling, and usage plans at the edge: predictable and extreme raceday spikes managed before they reach compute.
Lambda
Serverless compute behind API Gateway. Fastify adapted to Lambda for schema based serialisation of large numeric telemetry payloads.
Redis
Hot path caching for par times and pace figures, live in running state, idempotency keys for RabbitMQ consumers, and rate limiting at the API edge.
PostgreSQL + RDS
Tiered storage with date range partitioning across the full 25 year history. BRIN indexes and materialized views for expensive aggregates.
Lessons for enterprise, brands, scaleups & VC backed founders
What a 25 year data pipeline teaches you.
01
Entity resolution is the foundation
Everything downstream is only as trustworthy as the joins. Three providers with zero shared identifiers and inconsistent naming conventions made the resolution layer the single highest value component in the system.
02
Separate live from backfill
A 25 year historical replay pushes millions of messages. Separate queues with independent consumer pools and throttled prefetch ensure reprocessing never degrades live raceday: the design decision that matters most.
03
Smooth the signal, not the noise
Naive finite differencing of GPS position amplifies sensor noise into unusable velocity. Real pipelines need Kalman or Savitzky-Golay smoothing: the distinction between a real system and a demo.
04
Model data availability explicitly
The 25 year archive is deliberately mixed fidelity. Data availability is a first class, queryable attribute: consumers ask what is known about a race rather than discovering absence by surprise.
More case studies
Same team, more categories in production.
Polo Platform
USPA / HPA · AI sports analytics
AI powered match operations: video frame analysis, team and player analytics, predictive modelling, and MP4 to HLS pipeline with frame accurate seeking. Two governing bodies, one platform.
RYVYL
NASDAQ payments platform
Formerly GreenBox POS. Multi currency acquiring, real time settlement, treasury dashboards: engineered to a $43.5M NASDAQ IPO in 2021.
The Sufferfest
Consumer SaaS · Acquired
Napkin to acquisition consumer platform. $20M+ revenue, 70+ countries, 125,000+ community: acquired by Wahoo Fitness in 2019.
mydiveo
Forbes featured · Acquired
Mobile portfolio platform for artists. $4M raise, tier 1 press (Forbes, LA Times, Billboard), acquired for $7.4M within 15 months of launch.
SAFER
GovTech · Patented biometric
World's first biometric activated personal emergency alert system. Patented Jan 2017, beta tested with U.S. police departments.
REAP Pro
PropTech · Business Insider
Real estate appraisal platform replacing 70% manual data entry with OCR driven automation. Launched on Business Insider.
CaptureProof
AI health · Patented · $6.15M raised
AI driven visual health platform: Smart Medical Camera turning any smartphone into a clinical imaging device. 3 patents, 8+ awards including MedTech Breakthrough.
Saleem
HealthTech · 50K+ downloads · Iraq
Home healthcare marketplace for Iraq: doctor visits, nursing, physiotherapy, lab tests at your door. Arabic first, cross platform, Zain Great Idea Top Innovator.
Building a data pipeline, sports analytics, or IoT platform?
Ship the pipeline your analysts already trust.
30 minutes with the team that reconciles three live data providers, handles 25 year archives, and delivers telemetry grade analytics at raceday scale.