Artificial Intelligence 0725A005 · Final Project

SINAG vs Baron Crude

Can we predict the winner of an EV-vs-diesel race — and can we trust the prediction? A story about AI, real Philippine data, and what separates a trustworthy model from a lucky one.

The story · 1 of 2

The champion — and the shock

For years Baron Crude (diesel) won on raw power. Then the pump turned on him: NCR diesel ran ₱34 → ₱120.46/L through two war shocks — an all-time high in April 2026 (Strait of Hormuz). His monthly fuel bill became the whole story.CEIC

The story · 2 of 2

The challenger who doesn't burn a drop

Sinag — Tagalog for ray of sunlight — rolls in silent, charged at home overnight. His running cost barely moves: ₱1.46 → ₱2.45/km. The grandstands laughed… then the season started. Two questions decide it: can we predict the winner — and can we trust the prediction?

01 · Objective — what we predict

Predict real outcomes — and separate trust from illusion

We predict on real data: the EV-vs-diesel cost winner, and — on public course datasets — admission, diagnosis, species, sensor values. The deeper objective: show what makes a prediction trustworthy. Every claim faces three tests:

02 · Data

Sourced fuel data + six real ML datasets

NCR diesel · DOE/CEIC₱34→120.46
Meralco grid+75%
ML datasets6 · public

Fuel: NCR pump diesel + Meralco residential grid, 2020–2026, adversarially cross-checked & live-verified. ML: Iris, Breast Cancer (Wisconsin), UPCAT, Wave, Bike Sharing, Parkinson's — grounded in the course notebooks.refs

03 · AI concepts (technical)

kNN, linear classifiers — and the discipline

kNNPredict by the k nearest neighbours — vote (class) or average (value). k is the bias–variance dialCH67 LinearA weight vector = a boundary + margin (LogReg, LinearSVC); C = inverse regularizationCox58·CV95 DisciplineBeat a baseline (PCC / R²=0) · validate over repeated splits · avoid leakage
04 · Results & metrics

Strong where it should be — honest where it can't

ClassifyIris 96.97% · Breast Cancer 93.85% · UPCAT 86.9% (beats chance 63.4%) · Linear 95–96%grd RegressWave & Bike R² ≈ 0.76grd DieselR² −17.9 / −84.6 — a naïve forecast is worse than the mean. You can't trend-forecast a shock seriesdash
05 · Your turn — predict the winner

April 2026 · diesel at ₱120.46/L · 1,500 km

Before we reveal — call it out: is a home-charged EV cheaper to run than diesel this month? By how much per month?
06 · Impact & real-world value

Home-charged Sinag won every year — with an honest asterisk

For a PH driver or fleet, a home-charged EV beat diesel every year 2020–2026 (48–69%), up to 76% at the April peak — but only charging where you sleep; public DC erases it. The transferable win: the same discipline that crowns Sinag — beat the baseline, validate, no leakage — is what makes any AI prediction trustworthy.

Salamat! · Questions & Answers

Beat the baseline.
Validate. Trust it.

Appendix (swipe on) → model concepts (Ridge/Lasso, trees), the 50-question FAQ, and full references. Every number here is grounded in the course notebooks.

Reference material · for Q&A

Appendix

concept-only models · the 50-question FAQ · full references
— not part of the 7-minute talk —

Appendix · concept models not run here

Regularized regression — Ridge & Lasso

Ridge · L2Shrinks coefficients toward zero — tames multicollinearity & varianceHK70 Lasso · L1Shrinks and selects — some coefficients hit exactly zero (sparse)T96

No notebook was provided for this family — concept + reference only. The L1 diamond has corners on the axes, so its contour first touches there → a zeroed coefficient; the L2 circle has none.

Appendix · concept models not run here

Trees & ensembles

TreeRecursive splits → interpretable, but unpruned it overfitsCART84 Random ForestBagging + feature subsampling → variance ↓B01 Gradient BoostingSequential trees fit residuals under a lossF01

Also no notebook — concept + reference. Trees suit nonlinear, interaction-heavy tabular data — a different corner of the map from kNN or a linear boundary.

Appendix · FAQ 1/5

kNN, end to end

1 · What is kNN, how does k change it?
Lazy learner: store data, vote/average over k nearest. Small k overfits, large k underfits (Forge k=3 → 0.9143).
2 · Classification vs regression?
Same distance search: majority vote (Iris 0.9697) vs neighbor mean (Wave R² 0.7555).
3 · How is "best k" chosen honestly?
Scan k over repeated splits, report mean ± sd. Iris: k=11, 0.9697 ± 0.0208 over 20 splits.
4 · Accuracy on Iris & Breast Cancer?
Iris 0.9697 (k≈9); Wisconsin Breast Cancer 0.9385 (k=10). "Boston Cancer" is a misnomer.
5 · UPCAT — meaningful?
86.9% (k=38) clears 1.25×PCC = 63.40%, so it beats chance. (xlsx private → not reproducible.)
6 · Why does feature scaling matter?
Distance is its only similarity — large-scale features dominate. Standardize first (Bike 0.7636, k=5).
7 · Parkinson's k=1 = 0.9995?
Red flag: repeated-subject rows across a random split → 1-NN reads its own twin. Leakage, not performance.
8 · R² on Wave & Bike?
Both ~0.76: Wave 0.7555 (k=25, from 0.571), Bike 0.7636 (k=5).
9 · When does kNN break down?
High dimensions, irrelevant/unscaled features, large-n prediction cost; no compact model to ship.
10 · Sources?
Müller & Guido 2016; Fisher 1936; Fanaee-T & Gama 2014; Tsanas 2010; Cover & Hart 1967; Altman 1992.
Appendix · FAQ 2/5

Linear & regularized models

1 · Why is logistic "regression" classification?
Linear score → logistic link → P(class), thresholded to a label. Estimates a class, not a continuous target.
2 · LogReg vs LinearSVC?
Both linear boundaries; log-loss vs hinge (max-margin). On Cancer, within a couple points (0.9542–0.9636).
3 · What does C do?
Inverse regularization — larger C → weaker regularization → tighter fit, more overfit risk.
4 · Accuracy on Breast Cancer?
~95–96% across the C-sweep; a separate held-out report reaches 0.98 on 143 (confusion [[51,2],[1,89]]).
5 · What is it learning?
A weight vector + intercept; each coefficient is a feature's signed contribution (interpretable).
6 · Ridge (L2)? concept
L2 shrinks coefficients toward zero (none exactly zero) — tames multicollinearity/variance. Not run here.
7 · Lasso (L1) vs Ridge? concept
L1 drives some coefficients exactly to zero → feature selection, sparse model. Not run here.
8 · Feature engineering? concept
Derived (polynomial/interaction) features let a linear model fit curves. Not run here.
9 · When is linear the right call?
Roughly-linear signal, high dimensions, or when interpretability/speed matter.
10 · References?
Cox 1958; Cortes & Vapnik 1995; Hoerl & Kennard 1970; Tibshirani 1996; ESL 2009.
Appendix · FAQ 3/5

Trees, ensembles & validation

1 · How does a tree predict & overfit? concept
Axis-aligned splits to a leaf; unpruned it memorizes noise (high variance).
2 · How does Random Forest fix that? concept
Bagging + random feature subsets decorrelate trees → variance ↓.
3 · Gradient Boosting vs RF? concept
Sequential trees fitting residuals under a loss — lower bias, more tuning care.
4 · How to treat the tree claims?
Textbook expectations, not measured — no notebook, so concept + reference only.
5 · Choosing kNN vs linear vs trees?
Match data shape: scaling/low-dim (kNN), linear/interpretable (linear), nonlinear/interactions (trees).
6 · What is the PCC?
Chance accuracy given class prevalences; clear 1.25×PCC. UPCAT: 63.40%, and 86.9% clears it.
7 · Why repeated splits?
One split is a lucky draw; mean ± sd over many is honest. Iris k=11, 0.9697 ± 0.0208.
8 · What is data leakage?
Test info bleeding into training. Parkinson's k=1 → 0.9995 from repeated-subject rows.
9 · Connect to the diesel story?
Same discipline: the naïve OLS scored negative R² (−17.9/−84.6) — worse than the mean.
10 · How did each trial help?
Sweeping k shows bias–variance; the C-sweep maps regularization; k=1 exposes leakage. Comparison is the point.
Appendix · FAQ 4/5

Sources & methods

1 · Diesel price data?
Philippine DOE oil monitoring via CEIC — NCR pump advisories (₱34 → ₱120.46).
2 · Grid rates?
Meralco NCR residential rate advisories (delivered end-user rate).
3 · Brent & FX?
US EIA (Brent) and BSP (PHP–USD), backdrop only.
4 · ₱120.46 Apr'26 — verified?
Yes — live pass confirmed it against CEIC (120.458, an April 2026 record).
5 · ML datasets?
Fisher 1936; Wolberg et al.; Fanaee-T & Gama 2014; Tsanas 2010; Müller & Guido 2016; UPCAT private.
6 · Model results?
Transcribed verbatim from the MIB2026 course notebooks (Monterola), cached in docs/grounding/. Nothing re-fitted.
7 · Method citations?
Cover & Hart; Cox; Cortes & Vapnik; CART; Breiman; Friedman; Tibshirani; Hoerl & Kennard.
8 · Per-km assumptions?
12 km/L, 0.17 kWh/km, 1,500 km/mo. Shifting them moves the crossover; public DC can erase the edge.
9 · Exact vs approximate?
Event anchors exact; 2020–24 diesel annuals approximate; notebook metrics exact-as-run.
10 · Currency / period?
₱ PHP, nominal, 2020–2026, not inflation-adjusted.
Appendix · FAQ 5/5

Limits & what we can't claim

1 · Are Sinag & Baron real?
No — fictional characters; the pesos and gaps behind them are real and sourced.
2 · Did we fit ML to the diesel series?
No — grounded on the course datasets; diesel is the motivating story. Only a naïve OLS line, shown because it fails.
3 · Were Ridge/Lasso & trees run?
No — no notebook, so concept + reference, stamped not run here.
4 · Is UPCAT reproducible here?
No — the xlsx is private/not shared.
5 · Does this forecast diesel?
No — it characterizes a 2020–2026 regime; the OLS line scored negative R².
6 · Is Parkinson's 0.9995 real?
No — a leakage artifact, never a performance claim.
7 · Is "Boston Cancer" the housing set?
No — it's load_breast_cancer, the Wisconsin set. Misnomer corrected.
8 · Is the EV edge universal?
No — it depends on home charging; public fast-charging narrows or erases it.
9 · Does kNN accuracy transfer?
Not guaranteed — in-sample test splits of fixed benchmarks; mean±sd is split variance, not external validity.
10 · What can we legitimately claim?
Grounded, sourced results on named datasets + a sourced in-sample diesel-vs-Sinag comparison. Not forecasts.
Appendix · references 1/2

Data sources & datasets

Diesel data ✓ live-verified

  • NCR diesel ₱120.46 Apr'26 — CEIC ✓
  • Grid ₱14.83/kWh Jul'26 (+75%) — Meralco ✓
  • ₱153.70 single-station — GlobalPetrolPrices ~approx
  • EVIDA RA 11697 · TRAIN RA 10963 · excise-suspension RA 12316 (LawPhil) ✓
  • Brent EIA · FX BSP (backdrop) · OLS R² −17.9/−84.6 (dashboard.html)

ML datasets & notebooks

  • Iris — Fisher 1936 · Breast Cancer (Wisconsin) — Wolberg et al. 1993–95
  • Bike — Fanaee-T & Gama 2014 · Parkinson's — Tsanas et al. 2010
  • Forge / Wave — Müller & Guido 2016 · UPCAT — private
  • 10 course notebooks (Monterola, MIB2026) — docs/grounding/

Every ML figure is transcribed from the notebooks' printed outputs · gated by two adversarial checks (accuracy + anti-hallucination).

Appendix · references 2/2

Method citations

kNN & linear

  • Cover & Hart 1967 doi:10.1109/TIT.1967.1053964
  • Altman 1992 doi:10.1080/00031305.1992.10475879
  • Cox 1958 · Cortes & Vapnik 1995 doi:10.1007/BF00994018 · Fix & Hodges 1951

Regularized concept

  • Hoerl & Kennard 1970 · Tibshirani 1996 · ESL 2009

Trees & ensembles concept

  • CART 1984 · Quinlan 1986 · Breiman 2001 doi:10.1023/A:1010933404324 · Friedman 2001 doi:10.1214/aos/1013203451

Validation

  • Morrison 1969 (PCC) [verify] · Bergmeir & Benítez 2012

Full detail: docs/grounding/content-verified.md · Sinag & Baron Crude are fictional; the data & results are real and sourced.

SINAG·AI FINAL