feature_selection.md

Full Feature Set (SHAP Analysis)

SHAP Explanation for true_positive

Prediction Summary

• Baseline value: -0.0096
• Sum of SHAP values: 1.0267
• Final model output: 1.0171
• Output unit: model output

The final model output equals the baseline value plus the sum of all SHAP feature contributions.

Top Feature Drivers

Rank	Feature	Value	SHAP value	Direction	Impact share	Impact level
1	live_bets_count_log1p	3.1781	1.0110	increased the prediction	25.8%	medium
2	total_turnover	190.9000	-0.2828	decreased the prediction	7.2%	low
3	avg_deposit_amount_log1p	5.4650	0.2496	increased the prediction	6.4%	low
4	net_loss_rate	0.0715	-0.1843	decreased the prediction	4.7%	low
5	turnover_per_deposit	0.8114	0.1784	increased the prediction	4.5%	low
6	ggr_per_active_day	-3.0611	0.1707	increased the prediction	4.4%	low
7	avg_bet_amount	8.3000	-0.1529	decreased the prediction	3.9%	low
8	bet_day_rate	0.1000	-0.1304	decreased the prediction	3.3%	low
9	total_deposit_amount	235.2800	0.1280	increased the prediction	3.3%	low
10	avg_days_between_bets	3.3043	0.1250	increased the prediction	3.2%	low
11	ggr_margin_x_bet_day_rate	-0.0144	-0.1182	decreased the prediction	3.0%	low
12	registration_day_of_week	4.0000	-0.1083	decreased the prediction	2.8%	negligible
13	ggr_per_bet	-1.1978	-0.1021	decreased the prediction	2.6%	negligible
14	days_since_last_bet_x_bet_day_rate	0.8000	0.0962	increased the prediction	2.5%	negligible
15	active_to_tenure_ratio	0.0159	0.0829	increased the prediction	2.1%	negligible
16	payout_per_bet	9.4978	0.0595	increased the prediction	1.5%	negligible
17	ggr_positive	0.0000	0.0549	increased the prediction	1.4%	negligible
18	deposit_to_turnover_ratio	1.2325	0.0547	increased the prediction	1.4%	negligible
19	avg_bet_amount_x_bet_day_rate	0.8300	-0.0545	decreased the prediction	1.4%	negligible
20	first_bet_day_of_week	2.0000	0.0508	increased the prediction	1.3%	negligible
21	deposit_coverage_rate	0.9285	-0.0491	decreased the prediction	1.3%	negligible
22	activity_span_days	76.0000	-0.0485	decreased the prediction	1.2%	negligible
23	ggr_abs_log1p	3.3517	0.0442	increased the prediction	1.1%	negligible
24	ggr_margin	-0.1443	-0.0440	decreased the prediction	1.1%	negligible
25	bet_consistency_rate	0.1169	0.0431	increased the prediction	1.1%	negligible
26	avg_deposit_amount	235.2800	-0.0368	decreased the prediction	0.9%	negligible
27	turnover_per_active_day	21.2111	-0.0292	decreased the prediction	0.7%	negligible
28	net_loss	16.8300	-0.0243	decreased the prediction	0.6%	negligible
29	ggr_to_deposit_ratio	-0.1171	0.0241	increased the prediction	0.6%	negligible
30	last_bet_day_of_week	1.0000	0.0223	increased the prediction	0.6%	negligible

Plain-Language Summary

The strongest features pushing the prediction upward were: live_bets_count_log1p, avg_deposit_amount_log1p, turnover_per_deposit. The strongest features pushing the prediction downward were: total_turnover, net_loss_rate, avg_bet_amount.

Caveats

• SHAP values explain model behavior, not real-world causality.
• Correlated features can share or shift attribution.
• The meaning of a positive or negative SHAP value depends on the model target.
• These SHAP values are expressed in: model output.

SHAP Explanation for true_negative

Prediction Summary

• Baseline value: -0.0096
• Sum of SHAP values: -9.0680
• Final model output: -9.0776
• Output unit: model output

The final model output equals the baseline value plus the sum of all SHAP feature contributions.

Top Feature Drivers

Rank	Feature	Value	SHAP value	Direction	Impact share	Impact level
1	live_bets_count_log1p	4.9628	-3.9613	decreased the prediction	37.8%	high
2	avg_bet_amount_x_bet_day_rate	2.9377	-0.8320	decreased the prediction	7.9%	low
3	bets_per_active_day	4.5806	-0.6313	decreased the prediction	6.0%	low
4	tenure_x_bet_day_rate	161.1792	-0.4727	decreased the prediction	4.5%	low
5	ggr_margin_x_bet_day_rate	0.1489	-0.4388	decreased the prediction	4.2%	low
6	bets_per_tenure_day	0.3034	-0.3333	decreased the prediction	3.2%	low
7	ggr_per_active_day	16.8929	-0.3231	decreased the prediction	3.1%	low
8	ggr_per_bet	3.6879	-0.2571	decreased the prediction	2.5%	negligible
9	total_payout	687.5800	-0.2559	decreased the prediction	2.4%	negligible
10	avg_bet_amount	8.5300	-0.1921	decreased the prediction	1.8%	negligible
11	ggr_abs_log1p	6.2628	0.1903	increased the prediction	1.8%	negligible
12	payout_per_bet	4.8421	-0.1861	decreased the prediction	1.8%	negligible
13	last_bet_day_of_week	4.0000	-0.1802	decreased the prediction	1.7%	negligible
14	first_bet_day_of_week	5.0000	-0.1603	decreased the prediction	1.5%	negligible
15	deposit_to_turnover_ratio	0.6910	-0.1307	decreased the prediction	1.2%	negligible
16	turnover_per_deposit	1.4472	-0.1256	decreased the prediction	1.2%	negligible
17	days_since_last_bet	5.0000	-0.1201	decreased the prediction	1.1%	negligible
18	total_turnover	1211.2600	-0.1133	decreased the prediction	1.1%	negligible
19	days_since_last_bet_x_bet_day_rate	1.7220	0.1021	increased the prediction	1.0%	negligible
20	active_to_tenure_ratio	0.0662	-0.0979	decreased the prediction	0.9%	negligible
21	days_to_first_bet_in_obs	1.0000	-0.0965	decreased the prediction	0.9%	negligible
22	ggr_margin	0.4323	-0.0941	decreased the prediction	0.9%	negligible
23	ggr_positive	523.6800	-0.0935	decreased the prediction	0.9%	negligible
24	avg_days_between_bets	0.5845	-0.0885	decreased the prediction	0.8%	negligible
25	bet_consistency_rate	0.3690	-0.0798	decreased the prediction	0.8%	negligible
26	bet_day_rate	0.3444	0.0794	increased the prediction	0.8%	negligible
27	days_to_first_bet	380.0000	-0.0791	decreased the prediction	0.8%	negligible
28	total_deposit_amount	836.9500	0.0780	increased the prediction	0.7%	negligible
29	registration_month	9.0000	-0.0713	decreased the prediction	0.7%	negligible
30	ggr_to_deposit_ratio	0.6257	-0.0679	decreased the prediction	0.6%	negligible

Plain-Language Summary

The strongest features pushing the prediction upward were: ggr_abs_log1p, days_since_last_bet_x_bet_day_rate, bet_day_rate. The strongest features pushing the prediction downward were: live_bets_count_log1p, avg_bet_amount_x_bet_day_rate, bets_per_active_day.

Caveats

• SHAP values explain model behavior, not real-world causality.
• Correlated features can share or shift attribution.
• The meaning of a positive or negative SHAP value depends on the model target.
• These SHAP values are expressed in: model output.

SHAP Explanation for false_negative

Prediction Summary

• Baseline value: -0.0096
• Sum of SHAP values: -2.8281
• Final model output: -2.8377
• Output unit: model output

The final model output equals the baseline value plus the sum of all SHAP feature contributions.

Top Feature Drivers

Rank	Feature	Value	SHAP value	Direction	Impact share	Impact level
1	turnover_per_deposit	0.7888	-0.4970	decreased the prediction	12.7%	medium
2	total_deposit_amount	284.5600	-0.3627	decreased the prediction	9.3%	low
3	last_bet_day_of_week	6.0000	-0.2586	decreased the prediction	6.6%	low
4	days_since_last_bet_x_bet_day_rate	1.5560	-0.2439	decreased the prediction	6.2%	low
5	avg_deposit_to_avg_bet_ratio	27.2567	-0.2246	decreased the prediction	5.8%	low
6	ggr_margin_x_bet_day_rate	-0.0193	-0.2042	decreased the prediction	5.2%	low
7	ggr_per_active_day	-1.9929	0.1915	increased the prediction	4.9%	low
8	payout_per_bet	5.8688	-0.1578	decreased the prediction	4.0%	low
9	bet_consistency_rate	0.2059	-0.1436	decreased the prediction	3.7%	low
10	ggr	-27.9000	-0.1367	decreased the prediction	3.5%	low
11	deposit_per_bet	0.0465	-0.1227	decreased the prediction	3.1%	low
12	ggr_per_bet	-0.6488	-0.1215	decreased the prediction	3.1%	low
13	turnover_per_active_day	16.0329	-0.0988	decreased the prediction	2.5%	negligible
14	turnover_x_bet_day_rate	34.9260	-0.0976	decreased the prediction	2.5%	negligible
15	deposit_to_turnover_ratio	1.2678	-0.0863	decreased the prediction	2.2%	negligible
16	bets_per_active_day	3.0714	-0.0860	decreased the prediction	2.2%	negligible
17	avg_deposit_amount	142.2800	-0.0710	decreased the prediction	1.8%	negligible
18	total_payout	252.3600	0.0674	increased the prediction	1.7%	negligible
19	first_bet_day_of_week	2.0000	0.0634	increased the prediction	1.6%	negligible
20	days_to_first_bet	413.0000	-0.0594	decreased the prediction	1.5%	negligible
21	tenure_x_bet_day_rate	76.2440	-0.0594	decreased the prediction	1.5%	negligible
22	registration_day_of_week	2.0000	0.0565	increased the prediction	1.4%	negligible
23	net_loss_rate	0.1132	-0.0548	decreased the prediction	1.4%	negligible
24	days_to_first_bet_in_obs	12.0000	0.0547	increased the prediction	1.4%	negligible
25	registration_month	8.0000	-0.0453	decreased the prediction	1.2%	negligible
26	bets_per_tenure_day	0.0878	-0.0413	decreased the prediction	1.1%	negligible
27	ggr_to_deposit_ratio	-0.0980	0.0409	increased the prediction	1.0%	negligible
28	ggr_positive	0.0000	0.0406	increased the prediction	1.0%	negligible
29	avg_deposit_amount_log1p	4.9648	-0.0285	decreased the prediction	0.7%	negligible
30	total_turnover	224.4600	-0.0275	decreased the prediction	0.7%	negligible

Plain-Language Summary

The strongest features pushing the prediction upward were: ggr_per_active_day, total_payout, first_bet_day_of_week. The strongest features pushing the prediction downward were: turnover_per_deposit, total_deposit_amount, last_bet_day_of_week.

Caveats

• SHAP values explain model behavior, not real-world causality.
• Correlated features can share or shift attribution.
• The meaning of a positive or negative SHAP value depends on the model target.
• These SHAP values are expressed in: model output.

SHAP Explanation for false_positive

Prediction Summary

• Baseline value: -0.0096
• Sum of SHAP values: 0.6068
• Final model output: 0.5972
• Output unit: model output

The final model output equals the baseline value plus the sum of all SHAP feature contributions.

Top Feature Drivers

Rank	Feature	Value	SHAP value	Direction	Impact share	Impact level
1	live_bets_count_log1p	3.0910	1.1910	increased the prediction	22.0%	medium
2	turnover_x_bet_day_rate	9.5772	0.4695	increased the prediction	8.7%	low
3	first_bet_day_of_week	5.0000	-0.2362	decreased the prediction	4.4%	low
4	ggr_margin_x_bet_day_rate	0.0203	0.2257	increased the prediction	4.2%	low
5	activity_span_days	87.0000	-0.2228	decreased the prediction	4.1%	low
6	payout_per_bet	3.9600	-0.2132	decreased the prediction	3.9%	low
7	ggr_to_deposit_ratio	0.1313	-0.2049	decreased the prediction	3.8%	low
8	total_turnover	107.7300	-0.2041	decreased the prediction	3.8%	low
9	avg_days_between_bets	4.1429	0.1625	increased the prediction	3.0%	low
10	registration_day_of_week	4.0000	-0.1367	decreased the prediction	2.5%	negligible
11	total_payout	83.1600	-0.1269	decreased the prediction	2.3%	negligible
12	ggr_per_active_day	3.0713	0.1260	increased the prediction	2.3%	negligible
13	deposit_per_bet	0.0952	-0.1175	decreased the prediction	2.2%	negligible
14	ggr_margin	0.2281	-0.1136	decreased the prediction	2.1%	negligible
15	deposit_coverage_rate	0.4444	-0.1136	decreased the prediction	2.1%	negligible
16	avg_deposit_to_avg_bet_ratio	18.2398	0.1048	increased the prediction	1.9%	negligible
17	avg_bet_amount	5.1300	-0.1012	decreased the prediction	1.9%	negligible
18	bet_day_rate	0.0889	-0.0967	decreased the prediction	1.8%	negligible
19	bets_per_active_day	2.6250	-0.0944	decreased the prediction	1.7%	negligible
20	active_to_tenure_ratio	0.0235	-0.0941	decreased the prediction	1.7%	negligible
21	ggr_per_bet	1.1700	0.0937	increased the prediction	1.7%	negligible
22	tenure_x_bet_day_rate	30.3149	0.0850	increased the prediction	1.6%	negligible
23	avg_deposit_amount	93.5700	0.0796	increased the prediction	1.5%	negligible
24	net_loss	103.9800	-0.0775	decreased the prediction	1.4%	negligible
25	avg_bet_amount_x_bet_day_rate	0.4561	0.0712	increased the prediction	1.3%	negligible
26	bets_per_tenure_day	0.0616	0.0589	increased the prediction	1.1%	negligible
27	deposit_to_turnover_ratio	1.7371	0.0556	increased the prediction	1.0%	negligible
28	days_since_last_bet_x_bet_day_rate	0.0889	0.0494	increased the prediction	0.9%	negligible
29	ggr_positive	24.5700	0.0491	increased the prediction	0.9%	negligible
30	days_to_first_bet_in_obs	1.0000	-0.0455	decreased the prediction	0.8%	negligible

Plain-Language Summary

The strongest features pushing the prediction upward were: live_bets_count_log1p, turnover_x_bet_day_rate, ggr_margin_x_bet_day_rate. The strongest features pushing the prediction downward were: first_bet_day_of_week, activity_span_days, payout_per_bet.

Caveats

• SHAP values explain model behavior, not real-world causality.
• Correlated features can share or shift attribution.
• The meaning of a positive or negative SHAP value depends on the model target.
• These SHAP values are expressed in: model output.

Feature Selection Summary for Customer Churn Model

Assumption

This summary assumes the model’s positive class represents customer churn.

The SHAP analysis was reviewed across:

• True positives
• True negatives
• False positives
• False negatives

The goal is to identify which features should be kept, removed, or reviewed because they may be causing incorrect churn predictions.

---

Key Findings

The model relies mainly on:

• Betting activity
• Turnover
• Activity frequency
• Recency of betting behavior
• Deposit behavior
• GGR / profitability metrics

The strongest useful signals come from customer activity and engagement. However, several engineered ratio, interaction, and calendar-based features appear unstable and may be contributing to wrong predictions.

---

Features Recommended to Keep

These features appear to provide meaningful churn signal and should be kept for the next model iteration.

live_bets_count_log1p
total_turnover
turnover_per_active_day
avg_bet_amount
bet_day_rate
bets_per_active_day
bets_per_tenure_day
active_to_tenure_ratio
activity_span_days
avg_days_between_bets
days_since_last_bet
days_to_first_bet_in_obs
ggr_per_active_day
ggr_margin
ggr_abs_log1p
payout_per_bet
deposit_coverage_rate

Notes

• live_bets_count_log1p is the strongest feature overall, but it should be audited because it also contributes to false positives.
• Activity-based features such as bet_day_rate, bets_per_active_day, and active_to_tenure_ratio are useful because they capture customer engagement.
• Turnover and recency features are generally interpretable and should remain in the model.
• Only a small number of GGR features should be kept to avoid redundancy.

---

Features Likely Causing Wrong Predictions

These features appear to push the model in the wrong direction in false-positive or false-negative cases.

registration_day_of_week
first_bet_day_of_week
last_bet_day_of_week
registration_month
turnover_x_bet_day_rate
ggr_margin_x_bet_day_rate
avg_bet_amount_x_bet_day_rate
days_since_last_bet_x_bet_day_rate
tenure_x_bet_day_rate
turnover_per_deposit
avg_deposit_to_avg_bet_ratio
deposit_per_bet
deposit_to_turnover_ratio
ggr_to_deposit_ratio
ggr
ggr_per_bet
ggr_positive

---

Why These Features Are Risky

Calendar Features

registration_day_of_week
first_bet_day_of_week
last_bet_day_of_week
registration_month

These are likely weak or artificial signals. Raw weekday and month values can create misleading patterns because the model may treat them as ordered numbers.

Recommendation: remove them, or replace them with cyclical encodings only if there is a strong business reason.

---

Interaction Features

turnover_x_bet_day_rate
ggr_margin_x_bet_day_rate
avg_bet_amount_x_bet_day_rate
days_since_last_bet_x_bet_day_rate
tenure_x_bet_day_rate

These features appear unstable and contribute to incorrect predictions. They may be amplifying signals that are already captured by simpler base features. Recommendation: remove them and keep the cleaner base features instead.

---

Deposit Ratio Features

turnover_per_deposit
avg_deposit_to_avg_bet_ratio
deposit_per_bet
deposit_to_turnover_ratio
ggr_to_deposit_ratio

These features may cause the model to over-trust customers who deposited money, even when they later churned. This can increase false negatives. Recommendation: remove or reduce these features, especially if churn recall is important.

---

Redundant GGR Features

The model uses several overlapping GGR-related features. This can make the model harder to interpret and may split feature importance across similar variables. Recommended GGR features to keep:

ggr_per_active_day
ggr_margin
ggr_abs_log1p

GGR features to remove or validate carefully:

ggr
ggr_per_bet
ggr_positive
ggr_to_deposit_ratio
ggr_margin_x_bet_day_rate

---

Main Error Drivers

False Positives

False positives are customers predicted as churners even though they did not churn. Main risky drivers:

live_bets_count_log1p
turnover_x_bet_day_rate
ggr_margin_x_bet_day_rate
avg_days_between_bets
ggr_per_active_day
avg_deposit_to_avg_bet_ratio

The model may be overreacting to moderate activity gaps, certain live betting patterns, or engineered interaction features.

---

False Negatives

False negatives are customers who actually churned but were predicted as non-churners. Main risky drivers:

turnover_per_deposit
total_deposit_amount
last_bet_day_of_week
days_since_last_bet_x_bet_day_rate
avg_deposit_to_avg_bet_ratio
ggr_margin_x_bet_day_rate
payout_per_bet
deposit_per_bet
ggr_per_bet

The model may be treating deposit-heavy or previously active customers as safe, causing it to miss churn risk.

---

Recommended Next Model Experiment

Train the next model with a simpler feature set focused on:

• Activity frequency
• Turnover
• Recency
• Engagement over tenure
• A limited number of clean GGR features

Remove:

• Raw calendar features
• Handcrafted interaction features
• Deposit-heavy ratio features
• Redundant GGR features

This should make the model more stable, easier to explain, and less likely to produce wrong churn predictions caused by noisy engineered features.