Draft Pick Value

How to read it

The vertical axis is a value index, not a real-world unit — the 1st overall pick is pinned at 100 and everything else is a share of that. A pick at 50 has been, on average, worth half of a first-overall pick; a pick at 5 has been worth a twentieth. Hover any slot to read its value.

The chart shows three lines. The blue line is our model — the smoothed value curve. The amber line is the raw, unadjusted average for each individual pick: no compression, no smoothing, just what the players taken there actually went on to produce. It lurches up and down because a single late steal can carry a whole slot. The dashed green line is Dom Luszczyszyn’s published pick-value model (The Athletic, 2020), built on GSVA over a player’s first seven seasons — shown as a second opinion. His view is a touch flatter at the very top (he has the 2nd pick worth about 90% of the 1st, where ours says ~79%) and tracks ours closely through the rest of the draft. The dotted purple line is Eric Tulsky’s classic chart (Broad Street Hockey, 2013) — the GM-built curve fit to historical pick-trade behaviour; it drops far more steeply (2nd pick ~70% of the 1st) and treats day-three picks as nearly worthless.

Why it looks like this

Two things drive the shape. At the top, the first few picks routinely become franchise players (the McDavid / MacKinnon tier), and on raw WAR those players tower over everyone — so value falls off a cliff in the first handful of selections. In the tail, the opposite problem: roughly nine in ten players drafted in the late rounds never become NHL regulars at all, so the average career there is close to nothing and the curve flattens toward zero.

Two adjustments are baked into the curve. The very top is compressed a touch — a lone generational #1 like McDavid would otherwise make the 2nd pick look far weaker than it really is — and the late rounds are rolled smoothly toward zero to reflect how rarely they hit, without ever snapping to an artificial cliff.

The data & method

The outcome is career WAR (wins above replacement) from hockeystats.com. Every drafted skater from the 2013–2021 classes ( picks; goalies are excluded because WAR doesn't cover them) is matched to his career WAR by name; anyone who never reached the NHL counts as zero. For each draft slot we take the average across the nine classes, apply the two adjustments above, then average neighbouring slots together to steady the small sample (each exact slot has only about nine picks behind it), and pin 1st overall to 100. Values at a few representative slots:

The math, step by step

For anyone who wants the actual recipe: each draft slot starts as the list of career WAR values for the (up to nine) players ever drafted there, with anyone who never reached the NHL counting as 0. The blue line is then built in five steps.

1. Flatten the top. Before averaging, raise every player’s WAR to a power below one, w^0.64. A plain average lets one generational #1 (McDavid) tower over the field; the exponent pulls the giants back toward the pack. An exponent of 1 would just be the raw average.
2. Average each slot, blending neighbours. Each exact slot has only ~9 picks, so we take a weighted average across nearby slots with a bell curve, weight = exp(−0.5 · ((s−t)/h)^2) (s the slot, t its neighbours). The width h = max(0.7, 0.28 × slot) stays narrow at the top (so 1st and 2nd stay distinct) and widens down the board (so the thin late-round sample doesn’t read as a staircase). Busts stay in the count, so a slot full of misses is dragged toward zero.
3. Pin 1st overall to 100. Divide every slot by the first-overall value and multiply by 100 — turning WAR into a 0–100 index.
4. Never let it rise. A later pick shouldn’t be worth more than an earlier one, so we run isotonic regression (pool-adjacent-violators): wherever sample noise makes the line tick upward, that run is flattened to its average instead of stepping.
5. Roll the tail toward zero. Small values are eased down by a smooth gate, v · v^k / (v^k + c^k) with k = 2.5 and c ≈ 4.7 — it leaves big values almost untouched and pulls the late rounds toward 0 without ever snapping to a cliff. A final divide pins 1st overall back to 100.

Two numbers do all the shaping: the exponent a ≈ 0.64 (how hard to flatten the top) and the gate point c ≈ 4.7 (how hard to cut the tail). For comparison, the amber line skips every step above — it’s just the plain average WAR per pick, Σw / n, pinned to 1st = 100. The dashed green line is Dom’s, drawn by smoothly interpolating (monotone cubic) through the anchor values he published, on the same 0–100 scale.

Work in progress. This is a first build: names are matched loosely, so a few real NHLers get undercounted in the middle rounds, and with only about nine picks per exact slot the line is smoothed by averaging neighbouring picks. A production version uses cleaned name-matching and a leave-one-class-out fit. It also measures a player's career, not a draft pick's trade price: in the trade market (what teams actually pay in pick swaps) late picks change hands for far less than the careers here suggest — the very inefficiency that rewards trading down for extra picks.