Prediction System

[Tiapan1]

I have managed to create a system prediction system for this import of historical data and racecard data, has anyone else had any success in doing so?
if so, is it any good? have you tracked strike rate/ return on investment?

[odbball-a11y]

Hi Tiapan1 - I have and continue to try and build a UK Flat prediction system. My best effort so far is, training my model on about 3 years of data, then testing on about 1 year or 6k flat races.

The best i can do is, get to 22.5% on predicting winner of all 6k test races. If i make it selective, and we bet on approx half the races i can get to just over 28% win rate.

Odds fav wins 33%. To get to a positive ROI, i have to be very selective, so 6k races a year, i can bet on just over 360 races a year that return a 10% ROI.

...today i wanted to download more results for 2025, to see if my model could continue to show a profit....but Rpscrape keeps giving me an error trying to scrape the results, only spend about 30 mins trying to solve that ...is everyone else struggling to scrape results??

[warfishdata-cmyk]

Yes, definitely still struggling to scrape results, last week it would work periodically during off hours, but this week it almost always freezes up mid process. Likely blocked by RP. If anyone has a way let us know.

[Tiapan1]

yes this is the problem im currently facing......downloading the historical data right now, which is affecting the prediction system ive created, i have put a ticket in the issue tab,
it seems racing post have either changed the structure of their website or we are getting blocked by racing post/cloudflare, which are preventing downloads.
i have opened a ticket in the issue tab, however no has a solution for this at the moment.

[odbball-a11y]

What success have you had with your prediction system?
Can you turn a profit? Win %?

[Tiapan1]

i was currently tuning/tweaking my system, until the historical data files where unable to be downloaded.
the system was producing a win % of 21- 29%
Roi 60-70% when factoring higher odds/ value bets.

[odbball-a11y]

wow.....60-70% ROI......you have completed Horse racing betting.....i struggle to beat the overround!

[Tiapan1]

Would be nice if we had the data,
1 complete script to run off the back of historical data.
This has took me close to 2 years to produce.
These are the features I have included:

A. Core Rating / Form Features

Feature Meaning

or_rating Official Rating (OR) used in handicaps
rpr Racing Post Rating
or_scaled OR / 100
rpr_scaled RPR / 100
or_rpr_gap OR − RPR rating difference
expected_weight_from_or OR × 1.05 (expected lbs)
weight_diff_expected lbs − expected_weight_from_or

---

B. Weight / Draw / Age

Feature Meaning

lbs Carried weight
top_weight Max weight in race
weight_to_top top_weight − lbs
draw Stall draw
draw_bucket Draw quantile bucket
age Horse age

---

C. Distance Features

Feature Meaning

distance_f Distance in furlongs
distance_bucket sprint / middle / stayer / very_long

---

D. Positional & Recent Form

Feature Meaning

prev_pos_1 Previous finish position
prev_pos_2 2nd last position
prev_pos_3 3rd last position
last_pos prev_pos_1 fallback to 99
wins_last3 Wins in last 3
places_last3 Places (1–4) in last 3
avg_pos_3 Avg finishing pos over last 3

---

E. Days Since Run / Fitness

Feature Meaning

days_since_run DSR
dsr_log log(1 + DSR)
dsr_over30 1 if DSR > 30
dsr_over90 1 if DSR > 90

---

F. Sectional / Speed Figure Proxies

Feature Meaning

secs Finishing time (if provided in hist)
race_secs_mean Mean secs for that race
secs_minus_race_mean secs − race_mean

---

G. Pace (from comments)

Extracted from the text comment — 5 one-hot pace flags:

pace_led

pace_prominent

pace_mid

pace_held_up

pace_unknown

---

H. Improvement / Rating Momentum

Computed historically:

Feature Meaning

rpr_change_1 RPR today − last run
rpr_change_2 RPR today − two runs ago
or_change_1 OR today − last run
or_change_2 OR today − two runs ago
rpr_ma3 Rolling 3-run average RPR
or_ma3 Rolling 3-run average OR

---

I. Shape / Consistency of RPR/OR

Feature Meaning

rpr_diff RPR diff from last race
rpr_diff2 RPR diff from 2 races ago
or_diff OR diff from last race
or_diff2 OR diff from 2 races ago

---

J. Field Shape / Class Makeup

Feature Meaning

top_or Highest OR in race
bottom_or Lowest OR in race
or_spread top_or − bottom_or
or_std Standard deviation of OR in race

---

K. Going / Surface Interactions

Feature Meaning

surface_code Encoded surface (turf/aw/etc.)
going_heavy 1 if heavy
going_soft 1 if soft
going_firm 1 if firm
ground_speed Composite going-speed factor

---

L. Trainer & Jockey Exp-Decay Form (powerful)

Calculated for every race historically:

Trainer:

trainer_exp_win

trainer_exp_place

Jockey:

jockey_exp_win

jockey_exp_place

These measure last 180-ish days with exponential weighting.

---

M. Trainer & Jockey Time-Window Stats

For 14d, 30d, 90d windows:

For trainer:

trainer_runs_14d, trainer_win_14d, trainer_place_14d

trainer_rate_win_14d, trainer_rate_place_14d
(and same for 30d, 90d)

For jockey:

jockey_runs_14d, jockey_win_14d, jockey_place_14d

jockey_rate_win_14d, jockey_rate_place_14d
(and 30d, 90d)

---

N. Course-Level Features

From historical data:

course_win_rate

course_rpr_mean

course_field_mean

---

O. Course-Draw Bias Table

(merged where available)

course_draw_win_rate

[Tiapan1]

Bare in mind, this is for uk flat & jumps handicap horse racing only

[odbball-a11y]

I used the python code you posted...and it worked...so thank you ...i have managed to update my data.

[rmwesley99]

What exactly is your model trying to predict? On average, BSP is a very good predictor of a horse's chances of winning a race. It is a thankless task to try and get a better prediction than BSP using features crafted from publicly available data. ROI of 60% - 70% is not possible long term. It suggests your model is heavily backfitted to the data, or some future info has leaked into your features. You could try and find some edge cases where BSP is less efficient, but they tend to small margin (e.g. 1% to 4%) and transient as the markets shifts to become efficient to them. Other alternatives are to pick up pennies based on real time odds movement, either pre-off or in-running. For in-running lots use TPD data.

[Tiapan1]

The model is based on finding the predicted winner of a future race, based on historical data patterns, not BSP, none of my system is based on Odds at all.
Purely Stats.
i have stated above what the system is based on, and its around 90 features involved.
this isnt based on public opinions, which odds have factored in.

[Tiapan1]

and...what do you think ive been doing?
ive narrowed down to around 3 or 4 models.....
the models, try to find value, not just winners.

shap analysis and feature importance Done......
even told you im in the backtesting phase :)
everyone betting market favs, this is selecting value.

[rmwesley99]

Well done you! Never, ever reveal the secrets of how you got 70% ROI out of your backtesting though. Everyone will be doing it and there won't be any backtest profits left for the rest of us.

[Tiapan1]

why would i 👍
after all im guessing thats what you was hunting for?
someone to give you a quick fix for a perfect winning system, which will never exist ⭕
your looking for something a little more advanced that will pick you every winner for every race LMAO

[rmwesley99]

I've been rumbled. You are clearly a lot cleverer than you come across.

[Tiapan1]

👍