Page 1 Model Top-Pick Value Bets Three-year performance report All qualifying BSP prices | Out-of-sample test bets from 1 January 2023 to 31 December 2025 Prepared 21 July 2026 9,809 Qualifying bets +1,445.86 u Level-stakes profit 14.74% Gross ROI 141.46u Maximum daily-closing drawdown Overall finding: the 30% perceived-value strategy produced positive returns in every calendar year, every quarter and every race population. The daily-cluster bootstrap 95% interval for the three-year ROI is 8.8% to 20.8%. Executive findings Strong aggregate result: 9,809 bets returned 11,254.86 units from 9,809.00 staked, producing +1,445.86 units and 14.74% ROI. Time consistency: annual ROI was 12.6%, 17.2% and 14.4%. All 12 quarters were profitable and 28 of 36 months finished positive. Population breadth: every population was profitable in every year. Chase led with 26.6% ROI; hurdle was weakest at 6.9% and has the least conclusive confidence interval. Value filtering: compared with betting every model top pick, the 30% rule retained 34.2% of bets and 82.8% of profit, while increasing ROI from 6.10% to 14.74%. Calibration risk: actual equivalent wins were only 63.6% of model expectation, but 111.6% of market expectation. The selections beat BSP collectively, while the raw model probabilities overstated absolute win frequency. Robustness: removing the largest winning bet leaves 1,372.20 units and 13.99% ROI; removing the three largest leaves 1,290.20 units and 13.15% ROI. Model Top-Pick Value Bets - Three-Year Performance Report Page 2 1. Scope and methodology The analysis covers the supplied model_top_pick_value_30_plus strategy: one model top pick per eligible race when perceived value is at least 30% against the BSP-implied probability. All qualifying BSP prices are included. Only rows marked test are used, covering non-overlapping calendar-year test windows for 2023, 2024 and 2025. Perceived value: model probability / market probability - 1. Because market probability is 1 / BSP, the qualifying rule requires model probability to be at least 1.30 times the market-implied probability. Settlement: one-unit level stakes using the supplied returns and profit fields. Source metadata specifies gross BSP settlement, fractional dead-heat settlement and zero commission. The results therefore exclude exchange commission, liquidity constraints, price impact and rejected bets. Validation: bet-level stake, returns and profit reconcile to strategy-summary.csv, and the perceived-value formula reconciles to machine precision. strategy-daily(2).csv and strategy-summary.csv were used as cross-checks; all reported comparisons are rebuilt from strategy-bets.csv. Benchmark comparison Strategy Bets Stake Returns Profit ROI Strike Avg BSP All model top picks 28,653 28,653.00 30,399.66 1,746.66 6.10% 29.50% 5.07 Value >=30%, all prices 9,809 9,809.00 11,254.86 1,445.86 14.74% 17.94% 8.14 The value filter reduced turnover by 65.8% but surrendered only 17.2% of aggregate profit. This is the clearest evidence that the 30% signal concentrated the stronger portion of the model top-pick portfolio. Confidence intervals in this report use a non-parametric bootstrap that resamples whole settlement days rather than individual bets. This preserves same-day clustering but does not account for future model drift or changes in execution conditions. Model Top-Pick Value Bets - Three-Year Performance Report Page 3 2. Annual performance Year Bets Profit ROI Eq. wins Strike Avg BSP Max DD 95% ROI interval 2023 3,170 399.12 12.59% 542.5 17.11% 8.09 77.02 2.0% to 23.3% 2024 3,202 550.79 17.20% 584.0 18.24% 8.15 59.46 6.8% to 28.0% 2025 3,437 495.96 14.43% 633.5 18.43% 8.18 141.46 4.7% to 24.3% 2024 was the strongest year at +550.79 units and 17.2% ROI. 2023 was the lowest-ROI year at 12.6%, but its bootstrap interval remained above zero. 2025 produced +495.96 units despite the deepest annual drawdown. The annual results show persistence rather than dependence on one exceptional year: each period contributed materially to the three-year total, and no year was close to carrying the portfolio alone. Model Top-Pick Value Bets - Three-Year Performance Report Page 4 3. Profit path and drawdown 141.46u Maximum drawdown 03 Sep 2025 Prior equity peak 13 Oct 2025 Drawdown trough 65 days Peak-to-recovery The maximum daily-closing drawdown was 141.46 units. It began from the 03 September 2025 equity peak, reached its trough on 13 October 2025, and recovered on 07 Nov 2025. The historical peak-to-recovery period was 65 days. The curve advances in uneven steps rather than smoothly. Periods of flat or declining performance are normal for the observed strike rate and average price. The historical drawdown is descriptive, not a bound on future losses. Model Top-Pick Value Bets - Three-Year Performance Report Page 5 4. Quarter and month consistency Quarter Bets Profit ROI 2023Q1 492 112.41 22.8% 2023Q2 853 101.30 11.9% 2023Q3 982 135.22 13.8% 2023Q4 843 50.19 6.0% 2024Q1 487 47.87 9.8% 2024Q2 829 142.98 17.2% 2024Q3 990 214.22 21.6% 2024Q4 896 145.72 16.3% 2025Q1 506 84.47 16.7% 2025Q2 861 138.15 16.0% 2025Q3 1,028 53.77 5.2% 2025Q4 1,042 219.57 21.1% All 12 quarters were profitable. The strongest was 2025Q4 at +219.57 units; the weakest was 2024Q1 at +47.87 units. At monthly frequency, 28 of 36 months were profitable (77.8%). The best month was 2023-08 at +144.12 units. The worst was 2025-09 at -52.32 units. The monthly losses were more than offset by the breadth of profitable months and quarters. Model Top-Pick Value Bets - Three-Year Performance Report Page 6 5. Population comparison Population Bets Profit ROI Strike Avg BSP Max DD 95% ROI interval chase 1,627 432.24 26.6% 18.0% 8.54 52.33 10.0% to 44.4% flat 2yo 1,185 231.99 19.6% 21.9% 6.92 42.71 4.2% to 36.0% flat 3yo 1,340 206.05 15.4% 18.5% 7.68 65.33 0.3% to 31.1% flat 3yo plus 2,777 376.05 13.5% 15.9% 8.42 80.93 2.1% to 25.2% hurdle 2,880 199.54 6.9% 18.1% 8.37 109.65 -3.1% to 17.4% Chase was the strongest segment, generating 29.9% of total profit from 16.6% of bets. Flat 2yo ranked second on ROI and strengthened in 2025. Flat 3yo delivered modest returns in 2023-2024 before a much stronger 2025. Flat 3yo plus was a major profit contributor because of its scale, though ROI softened in 2025. Hurdle remained profitable but delivered the weakest ROI and the only population-level confidence interval that includes zero. Model Top-Pick Value Bets - Three-Year Performance Report Page 7 6. Population stability and contribution All 15 population-year combinations were profitable. That breadth is a stronger result than the aggregate population ranking alone, because no segment depends on a single profitable year to offset a loss elsewhere. Chase generated 29.9% of profit from 16.6% of bets. Hurdle represented 29.4% of bets but only 13.8% of profit. Flat 3yo plus contributed close to its volume share, making it an important scale segment even without the highest ROI. Model Top-Pick Value Bets - Three-Year Performance Report Page 8 7. Population profit paths and calibration The population curves show different return patterns. Chase produced the steepest sustained growth; flat 3yo and flat 2yo accelerated late in the sample; flat 3yo plus was strong early but flatter through parts of 2025; hurdle experienced the longest recovery period. Every population produced fewer equivalent wins than the model probabilities predicted, with actual/model ratios between 60.6% and 67.6%. Every population nevertheless beat its BSP-implied expected wins, with actual/market ratios between 107.2% and 118.3%. This combination indicates useful ranking or value discrimination against the market, alongside systematic overstatement of absolute win probabilities. Model Top-Pick Value Bets - Three-Year Performance Report Page 9 8. Robustness and conclusions Robustness to the largest winners Scenario Profit ROI Full strategy 1,445.86 14.74% Excluding largest winner 1,372.20 13.99% Excluding three largest winners 1,290.20 13.15% The strategy remains strongly profitable after removing the largest winning observations. This reduces concern that the headline result is an artefact of one or two exceptional payouts. The same check remains positive within every population, although hurdle is naturally the least robust because its aggregate margin is smaller. Conclusion The evidence supports the 30% perceived-value strategy as a broad, repeatable historical edge across the supplied 2023-2025 out-of-sample data. Profitability was positive across years, quarters and populations, while the result remained substantial after removing the largest winners. - Primary strength: consistency across independent time periods and race populations, rather than a single concentrated source of profit. - Best population: chase combined the highest ROI with a positive result in all three years and a confidence interval comfortably above zero. - Most important monitoring population: hurdle, because it has the lowest ROI, longest historical recovery and the widest uncertainty relative to its edge. - Main model risk: probability overconfidence. Actual win frequency consistently fell below the model expectation despite exceeding the market expectation. - Deployment caveat: reported returns are gross BSP. Commission, execution, liquidity and model drift can materially reduce live performance. - Recommended monitoring: continue tracking annual and rolling-quarter ROI, population contribution, drawdown and actual-to-model versus actual-to-market expected wins. Limitations - The report evaluates historical out-of-sample rows only; it does not guarantee future profitability. - Gross BSP and zero commission overstate net exchange returns where commission or execution slippage applies. - No liquidity, minimum-stake, rejected-bet or market-impact model is included. - Bootstrap intervals capture sampling variation at the daily level but not structural model changes or future market adaptation. - Population comparisons are descriptive; differences in race mix, average BSP and bet volume affect direct ROI comparisons. Model Top-Pick Value Bets - Three-Year Performance Report Page 10 Appendix A. Population detail Population Bets Eq. wins Strike Profit ROI Mean value Actual/mod el Actual/mark et Positive months chase 1,627 292.5 17.98% 432.24 26.57% 104.9% 66.1% 118.3% 24/36 flat 2yo 1,185 259.0 21.86% 231.99 19.58% 92.5% 67.6% 117.7% 16/30 flat 3yo 1,340 248.0 18.51% 206.05 15.38% 98.0% 64.4% 114.5% 19/34 flat 3yo plus 2,777 440.5 15.86% 376.05 13.54% 86.2% 63.1% 107.2% 22/36 hurdle 2,880 520.0 18.06% 199.54 6.93% 111.1% 60.6% 107.8% 18/36 Appendix B. Data provenance File Rows Use in report strategy-bets.csv 76,861 Primary bet-level analysis strategy-daily(2).csv 15,798 Daily-series validation strategy-summary.csv 60 Aggregate reconciliation and settlement metadata All calculations are reproducible from the uploaded files. Monetary values are stated in betting units because no currency denomination was supplied. Equivalent wins include fractional dead-heat outcomes.