Hill Race Families

1 What kinds of hill race are there?

Every race here rewards HL — that is what puts it on this list. What separates them is the second stat, and that is what decides who actually wins. Rather than assume the categories, this groups the 30 hilly races in division 3 (HL weight ≥ 6) by the shape of their own weightings, and lets the data choose how many groups there are.

These are slices of a continuum, not natural groups. The best silhouette score across every k tested is only 0.251 — well below the ~0.5 that indicates real cluster structure, and it barely moves as k changes. The hill calendar varies smoothly between its extremes rather than falling into distinct types. Treat the families below as useful labels for regions of that space, not as hard categories.

How many families?

Silhouette rewards tight, well-separated groups (1.0 is perfect; 0 means they overlap completely). Within-SS always falls as k rises, so it cannot choose k on its own.

The two axes that actually matter

• PC1 — 38.5% of variance: CL +0.54 DH +0.44 HL +0.20 SP -0.40 FR -0.50 CB +0.00 TQ -0.24 TT -0.04
• PC2 — 29.4% of variance: CL +0.06 DH +0.27 HL -0.82 SP -0.02 FR -0.21 CB +0.19 TQ +0.39 TT +0.15

Each dot is one race, placed by its two main components of variation and coloured by family. The spread is even, with no gaps between groups — that is the continuum showing through. The first axis is the real story: races trade climbing weight against sprint weight, smoothly.

The families

2 Who dominates each type?

The top five by average finishing place in each family, across seasons 78–79 in division 3. Read the stat columns down the page: notice how few of these riders win on one elite stat. Almost all of them carry a strong second and third.

Ranked by average place among riders with starts in at least half a family's races. Average place rewards consistency, so a rider who skips the hard editions flatters himself — read it alongside the start count.

3 Is there a tour for hillers?

A hiller's tour and a climber's tour are not the same thing. The first loads HL; the second loads CL/DH. Ranking every division 3 tour by the share of its total stage weight sitting in each:

HL   CL+DH   SP+FR+TQ. "Hilly" counts stages with HL weight ≥ 6.

Tour of the Western Cape is the hillers' tour — 29% of its weight is HL, the highest in the division, with 3 of 4 stages at HL ≥ 6 and a mean HL weight of 6.5. It is not the same as the climbers' tour: Copa de la Historia de los Andes carries the most CL+DH at 44%, but only 18% HL.

The catch: Tour of the Western Cape is also 59% sprint weight. A pure hiller gets in the door on HL, then loses the finish to SP/FR. It rewards a hiller who can sprint, not a hiller alone.

Tour of the Western Cape — stage by stage

General classification — season 78

10 of the top 10 on GC are hillers (HL ≥ 70). The overall win is decided on the hills, not on the flat.

Stage winners — season 78

Every stage went to a hiller.

4 Training proposal — Arjan Woering

dodge  ·  age 26  ·  CL 35 DH 40 HL 95 SP 30 FR 30 CB 27 TQ 34 TT 27

What is left to spend

71 of the 180-point lifetime cap is spent, leaving about 109. Almost all of it went into one stat. HL is effectively finished — the label still reads Incredible, but the growth has flattened, and the last few points there are the most expensive on the board.

What each stat pays, per point

Room is what the ceiling allows; rate is the mean-rank improvement per training point if that stat is trained alone from here. A stat can be the best buy per point and still not be the biggest total gain, because it runs out of room.

Calendar weight is the sum of that stat's convex race weight across all hill races — the reason the rates come out where they do. CB and TT are effectively dead weight on this calendar.

Where those 109 points are worth most

Each option below spends the same budget a different way, then ranks the resulting rider against all 1356 riders in the division, race by race. Lower is better.

Mean rank is across all hill races; the family columns break the same figure down by type. A plan can only spend what its stats can absorb — where "Pts" falls short of 109, the chosen stats hit their ceiling first.

The recommendation

Spread all five — 109 points, taking him to CL 55 DH 67 SP 56 FR 45 TQ 55. Mean rank 27 against 399 untrained, with 17 top-ten finishes across the hill calendar instead of 0.

The single-stat plans all lose, and they lose for two different reasons. Piling everything into one secondary leaves him a passenger in every family that does not weight it — and the ceilings bite: a "Small" potential stat runs out of room long before the budget does, stranding points he could have spent elsewhere.

Where he has actually been raced

Worth comparing against the family columns above before spending anything — a schedule change costs nothing and takes effect immediately.

Check the ceilings before committing. Potential bands are not in the database; they come off the in-game profile, and the true maximum inside a band is random. The ceilings assumed here are CL 55, DH 75, SP 65, FR 45, CB 55, TQ 55, TT 45. Train the first stat and watch where it stalls — that resolves the rest of the plan.

Method

Race weights (0–9 per stat) are converted through the game's raceStat curve, which is strongly convex — a weight of 8 is worth 22.0 against 4.5 for a weight of 3 — then normalised to each race's share of total weight, so races are compared by shape rather than size. TT keeps the 1.2× bonus it has in the race score. Families come from k‑means over those compositions, with k chosen by silhouette. The training simulation scores a hypothetical rider with the same formula against the live division field; it captures stat fit only, not tactics, form, DP/RS or team support, so treat the ranks as relative rather than as predicted finishes.