TennisEdge's AI model finds no betting value in T. Fritz vs Z. Bergs (Washington on 2026-07-27) — the market is efficient, with T. Fritz favoured at 1.23.
═══ T. Fritz vs Z. Bergs — Washington (hard) — 1/16-finals ═══
MODELS vs MARKET
Market (Betfair ml): Fritz 1.26 (79.4%) / Bergs 4.70 (21.3%); book 1.23/4.10.
ELO: Fritz 0.716 | ML: Fritz 0.602 | blend: Fritz 0.684. value_pick B, edge 8.5%.
Sim baseline: Fritz 0.634, exp margin 1.7, spw_a 0.601.
Sim re-run (spw_a 0.601->0.651, +0.05 clamp, justified below): Fritz 0.840, margin 4.3.
ML (0.60) is the outlier dragging the blend under market; ELO (0.72) sits near market.
INPUT AUDIT (resimulate) — the crux
Fritz spw 0.601 is anomalously LOW for one of the tour's biggest servers. His
30-match sample is entirely grass/clay (Wimbledon, Halle, Stuttgart, RG, Geneva)
against an elite-heavy field (Zverev x2, Shelton x2, Bublik x2, Tiafoe) — that
deflates his serve/return averages, and grass/clay numbers do not represent his
HARD-court serving. This match is on hard, Fritz's best surface. A bounded +0.05
shade (to 0.651, still conservative) lifts him to 0.840 win / 4.3-game margin. The
sim is knife-edge sensitive around this input: 0.601->0.651 swings the win prob by
20 pts and Bergs +4.5 cover from 0.632 to 0.396. Conclusion: baseline UNDER-rates
Fritz; the "Bergs value" is a grass-input artifact, and market ~0.79 (bracketed by
0.63 and 0.84) is the credible number.
RECENT FORM (last 5)
Fritz (A): Lost Zverev 1913/Elite 4-6 4-6 2-6 (straights, blowout finish); Won
Bublik 1856/Elite 7-6 6-4 6-4 GRIND; Won Sonego 1675/Strong 4-6 6-3 6-4 7-6 GRIND;
Won Kypson 1550/Avg 6-2 6-2 7-5; Won Lajovic 1524/Weak 6-3 6-4 6-3 DOM. Wimbledon
QF, Halle final, beat Zverev+Shelton on grass. Elite hard-court player. Shape STABLE
(QF loss to elite Zverev in straights = neutral).
Bergs (B): Lost Fery 1706/Strong (5 sets); Won Faria 1727/Strong 7-6 4-6 6-2 6-3;
Won Humbert 1764/Strong (5 sets); Won EASTBOURNE FINAL over Humbert 1794 3-6 6-1 6-4
(TITLE 06-28, grass); Won Samuel 1609/Avg. Grass-title form in June, but NO hard-court
matches in sample; clay results earlier were poor. Shape STABLE/slightly FADING (early
Wimbledon exit to lower-ranked Fery after the title).
H2H: Fritz beat Bergs 06-17 Halle (GRASS), 3 tight sets 6-7 7-5 4-6 — close, deciding
set, Bergs within ~2-3 games. Fritz won on Bergs' relatively better surface.
FATIGUE / PATH: fatigue null-equivalent, both well_rested (zeros). Last matches Fritz
07-08 (19d), Bergs 07-04 (23d) -> RUST risk both; both first match at Washington.
Grass->HARD TRANSITION for both (shared) — widens margin uncertainty, argues against
any tight/wide cover thesis. Ages 28/27, baseline recovery, no fatigue penalty anyway.
MARKET-BY-MARKET
Bergs ML @4.70: baseline edge is real numerically, but the input audit shows it comes
from a deflated Fritz serve number; corrected sim erases it (Fritz 0.84). PASS.
Fritz ML @1.26: effective implied ~0.80; even at corrected 0.84 edge <5% and 0.84 is
the top of my clamp — no value. PASS.
Bergs +4.5 @1.84: cover swings 0.632->0.396 on a 0.05 input change; straddles the
~0.556 (commission-adj) break-even; high-variance line (|line|>=4.5, needs 10%). No
stable edge. PASS.
Fritz -4.5 @2.10: corrected cover 0.604 -> looks like ~11% edge, but ONLY at the top of
the clamp; baseline cover 0.368. High-variance (laying fav wider than -2.5). See gates.
Totals 22.0: baseline exp 22.3 / corrected 20.4 — input-sensitive, no stable edge. PASS.
GATES (best candidates: Bergs ML value-pick, and Fritz -4.5)
G1 edge-origin: FAIL — the Bergs ML edge exists only because sim/ML inputs deflate
Fritz's hard-court serve (grass/clay sample); it is not a robust model-vs-market gap.
Any Fritz -4.5 edge appears solely at the top of the +/-0.05 clamp, not robustly.
G2 consensus: FAIL — market (Fritz 79%) sits against Bergs; the only support is a
grass-form model number the input audit discredits. No on-HARD H2H or live-hard
evidence for a Bergs upset; the lone H2H (grass) went to Fritz.
G3 fatal-risk: FAIL — the one H2H was a TIGHT 3-setter (Bergs within ~2-3 games) and
Bergs is a grinder who stays close; that directly negates a Fritz -4.5 blowout cover,
while the surface reset to hard is exactly where the models are least reliable.
VERDICT: PASS. Classic model artifact — ML/baseline sim under-rate an elite hard-court
server off a grass/clay sample, manufacturing "Bergs value" the market rightly ignores.
Corrected sim brackets the market price; no market offers a robust, gate-clearing edge.