Hong Kong Speed Figures

[La Cressonnière]

Hi,
Using the same approach, here is the selection of bets for tomorrow at Happy Valley. The idea is to select the correct line corresponding to the horse — preferably the most recent one, unless the horse was impeded — on the same racecourse and at the same distance. Here is the betting selection for tomorrow.

 

[Outlander]

Wednesday HV

Hi,
Using the same approach, here is the selection of bets for tomorrow at Happy Valley. The idea is to select the correct line corresponding to the horse — preferably the most recent one, unless the horse was impeded — on the same racecourse and at the same distance. Here is the betting selection for tomorrow.

pretty good start there La Cressonnière , plenty of them placed, and a few winners

 

[La Cressonnière]

Wednesday HV

pretty good start there La Cressonnière , plenty of them placed, and a few winners

Thanks O Outlander
22 selections
4 winners = 34,5
11 place = 29,9
3 quinella win = 33,1
6 quinella place = 202,25
1 Trio = 21,7
1 first 4 = 30,5
(HK DIVIDENDS) 1 unit

 

[tacker]

I have little idea whether or not TPD is even close to being accurate or not but unless there's something on offer that is superior then i've no alternative but to take it as gospel.
My interest in sectionals is derived from reading form and recognising the anomalies that crop up and distort the logic.
Official ratings and the final times ought to be something that we might be able to rely on but because we see the sectionals now it has become clear they don't and like we've discussed many times O Outlander it is down to the early pace, whether this be too fast or more likely too slow.
If we take the two races mentioned from the wolves meeting and repeat the points already made those two races undermine the "form" and if trying to put numbers to the final times are to mean anything then we need to find some logic behind them.

18:00 | Wolverhampton | Tuesday 28th October 2025 | At The Races

18:00 | Wolverhampton | Tuesday 28th October 2025 | Always Gamble Responsibly At BetMGM Fillies' Handicap

www.attheraces.com

17:00 | Wolverhampton | Tuesday 28th October 2025 | At The Races

17:00 | Wolverhampton | Tuesday 28th October 2025 | Join The Midnite Movement Apprentice Handicap

www.attheraces.com

A cl5 race & cl6 with ratings to match but the final times don't work out as we might think, this is purely down to the early pace of the cl5 race being too slow as demonstrated by the numbers.

At the end of the day as markfinn says it needs to go somewhere, while it is fairly easy to take the numbers into account from a form point of view that won't help if you want the numbers to tell you where to go, hopefully something can emerge but maybe we're on the wrong thread.

Sorry to trouble this thread again but just wanted to show some very different numbers from MARIS ANGEL yesterday.

15:22 | Wolverhampton | Tuesday 4th November 2025 | At The Races

15:22 | Wolverhampton | Tuesday 4th November 2025 | Make The Move To Midnite Handicap

www.attheraces.com

 

[Outlander]

T tacker
more interesting uplifts if you count in 2f sections, think more sensible, uplifts your Maris Angel race much more
i use 400m sections in HK , maybe using every furlong minimises the efficiency uplifts too much
View attachment 161642

View attachment 161643

View attachment 161644

Interesting T tacker that the horse was within 0.06 seconds of the uplifted time I got when I used 400m sections, probably means nothing.

 

[tacker]

Interesting T tacker that the horse was within 0.06 seconds of the uplifted time I got when I used 400m sections, probably means nothing.

To be honest O Outlander you will know better than i do and maybe you shouldn't be too quick to dismiss it though not sure how you can test it further.

 

[AustinDillon75]

Hi,
Here is what I’m trying to do.
XGBOOST MODEL
Based on the results from HK Happy Valley – September 2024.
A total of 3,766 performances were used.
Race distances included: 1000m, 1200m, 1650m.

The objective of the model is to predict which horse can finish in the top three, using the features from the Outlander dataset as well as additional variables, which we will discuss later.

The model’s metrics at Happy Valley are strong.
I carried out an initial test (TEST SET) on the races held at Happy Valley since September 10th, 2025 — a total of 21 races.
The model suggested 31 bets to place (prediction output > 0.45).
26 were successful, meaning the horses finished in the top three → ROI: 199%.
A file containing these results and the corresponding payouts is attached.

Good luck with that mate, sounds really promising. The one thing I have found with stuff like this is if you aren't careful you can find that a model might over-fit, or be prone to data leakage. What I mean by the latter is the model could be inadvertently "peeking" at the day's actual results and it can inflate the true performance to completely unrealistic levels. Not saying it's happened with those results but it's just something to watch for. Are you applying a train, test and validate approach, if I might ask?

 

[La Cressonnière]

Good luck with that mate, sounds really promising. The one thing I have found with stuff like this is if you aren't careful you can find that a model might over-fit, or be prone to data leakage. What I mean by the latter is the model could be inadvertently "peeking" at the day's actual results and it can inflate the true performance to completely unrealistic levels. Not saying it's happened with those results but it's just something to watch for. Are you applying a train, test and validate approach, if I might ask?

 

[La Cressonnière]

Hi
Yes, I validated all the optimization parameters after a 5-fold cross-validation by applying them to the entire dataset — no overfitting. Now i am working on SHA TIN model.

 

[The Hare]

Great work La Cressonnière

Here are your results and card for Sunday. Another Group 3 handicap. Hope you have your model complete in time. I am still working on my time ratings, hopefully time next couple of days.

 

[La Cressonnière]

Hi,
The SHA TIN model is almost finished. 10 features. It will be active next week. I'm starting with 1 training set + 1 test set + 1 blank data set for validation of unknown performance. The results at this stage are very encouraging before moving into production.
AustinDillon75
XGBoost Binary Image Classification Model – Final Report
After **5-fold cross-validation with early stopping**, the model achieves:
- **Test AUC: 0.971**
- **Virgin (unseen) Data AUC: 0.968**
- **F1-score on virgin data: 0.827**
- **Cohen’s Kappa: 0.770**
No overfitting. Excellent generalization. Production-ready
Model Configuration**
| Component | Details |
|---------|--------|
| **Algorithm** | XGBoost Classifier |
| **Features** | 10 (including `SPDNW`, `400NW`, `SPDNW1`, `400NW1`) |
| **Target** | Binary (0/1) |
| **Training** | 5-fold CV + **Early Stopping** (patience=50, eval_metric='auc') |
| **Best CV Score** | **0.8998** (mean AUC across folds) |
| **Best Iteration** | 61,920 (boosting rounds) |
## **3. Performance Metrics**
| Dataset | Accuracy | AUC | F1 | Recall | Precision | Kappa | FPR |
|--------|----------|-----|-----|--------|-----------|-------|-----|
| **Train** | 0.968 | **0.998** | 0.936 | 0.997 | 0.883 | 0.915 | 0.041 |
| **Test** | 0.899 | **0.971** | 0.801 | 0.863 | 0.747 | 0.734 | 0.090 |
| **Virgin Set** (unseen) | **0.914** | **0.968** | **0.827** | **0.887** | **0.775** | **0.770** | **0.078** |

> **ΔAUC (Train → Virgin) = 0.030** → **Minimal overfitting**
## **4. Confusion Matrix – Virgin Set**
| | Predicted 0 | Predicted 1 |
|---|-------------|-------------|
| **Actual 0** | 189 (TN) | 16 (FP) |
| **Actual 1** | 7 (FN) | **55 (TP)** |
- **Only 7 critical misses** (FN)
- **High precision-recall balance**

 

[The Hare]

Results and card for Wednesday

 

[La Cressonnière]

HV 12/11/2025
I removed the SP variable from the learner. I find that it's too closely tied to the market price in the forecast. My next idea is to include Benter's calibrated probability. If anyone has used it before, I'd be interested in their feedback. PROB > 0.65

R1	R2	R3	R4	R5	R6	R7	R9
DRACO	TALENTS SUPREMO	CALIFORNIA BANNER	SETANTA	ROBOT LUCKY STAR	LOOKS OUTSTANDING	FIND MY LOVE	FALLON
APOLAR FIGHTER		LESLIE	TAKE ACTION	STAR FIGURE	ISLAND BUDDY	FUN ELITE	SOLEIL FIGHTER
PERFECTO MOMENTS		AEROINVINCIBLE	QUANTUM PATCH	KING GLORIOSO		KING MILES	HUGE WAVE
		REWARDING TWINKLE
		BULL ATTITUDE

 

[The Hare]

Excellent La Cressonnière .

I think you are well ahead of me in knowledge. I know adding the SP or current price was what turned Bentner's model into profit. But I don't know his calibrated probability. Do you mean the weighting he gave to each input? Don't know if that is out there anywhere. Be interesting if it is.

Excellent work.