AuctionIQ

Prediction vs reality: AuctionIQ's top-priced player and likeliest buyer against the real 2026 auction
A price call: Venkatesh Iyer, AuctionIQ median ₹8.50 Cr, sold for ₹7.00 Cr
Held-out test on the whole real auction, run before any change to the model
A live lot in the 3D auction room, every number sealed until the hammer
The hammer falls: this run set against all 1,000 runs
Why this price? Worth, demand and how the bidding ended
KKR's table analyst explains, in cricket terms, why they are in

Inspiration

Every December, ten IPL franchises walk into one room with one purse each and seconds to decide every bid. Overpay early and you're broke later; hesitate and the player is gone. We wanted to see that room before it happens.

What it does

AuctionIQ predicts the IPL 2026 mini-auction (16 December 2025, Abu Dhabi) before it happens: who buys who, and for how much. It re-runs the whole auction 1,000 times using only what was known that morning, then lets you watch any lot play live in a 3D auction room.

Live lots, sealed until the hammer. Pick a player and the app fast-forwards to his lot and plays one of the 1,000 runs bid by bid. Every number stays sealed until the hammer falls; only then does it set the price against all 1,000 runs (the median, and who bought him how often).
Table analysts. Tap any franchise table and its analyst explains, in cricket terms, why that team is in or why it walked away: squad gaps, purse, overseas slots.
Why this price? Every sale explains the player's worth, the demand in the room and how the bidding ended.
Team habits and search. Open any franchise to see its gaps and buying habits, or search all 369 players. Works on a laptop or a phone.

How we built it

Player ratings. Ball-by-ball data from about 5,000 men's T20 matches (Cricsheet) becomes IPL-equivalent skills per player: runs added per ball against the same phase and venue, adjusted for league strength, weighted toward recent seasons and shrunk by evidence, with roles from usage and matchups against six bowler types.
Value model. A censored regression trained only on earlier auctions of the same kind (mini or mega) turns skills, caps, age, base price, previous price, the player's place in the official list and pre-auction buzz for uncapped players into a price.
Ten franchise agents. Each rule-based agent has its real purse, slots and retained squad. It opens a bid gate only for players it needs and bids up to a walk-away price it can explain: value × need × scarcity × purse pressure × style × noise.
Auction engine. Follows the published procedure (sets, a random draw inside sets, the price-point bid ladder, accelerated rounds, purse reserve, overseas fee cap) and runs it 1,000 times (Monte Carlo) in Python with NumPy and pandas.
App. A static web app (Vite + three.js) with a 3D auction room, deployed on Vercel and GitHub Pages.

How well it works

Held-out test. Before any change to the model, a separate script compared the simulation with the real 2026 auction. It separated the players who sold from those who did not far better than base prices (AUC 0.81 vs 0.62), and had the real buyer among its three likeliest teams 45% of the time (30% by chance). The model then gained a pre-auction buzz signal for uncapped players, re-tuned on the 2024 auction only; that is the version in the app.

Calls that came true (in both versions):

AuctionIQ's most expensive player was Cameron Green, most likely to KKR. In the real auction Green was the top buy: KKR, ₹25.20 Cr.
The likeliest buyer was the real buyer for Wanindu Hasaranga (LSG) and Sarfaraz Khan (CSK).
A price call: Venkatesh Iyer, base ₹2.00 Cr. AuctionIQ's median was ₹8.50 Cr in the app (₹8.00 Cr in the held-out run); he sold to RCB for ₹7.00 Cr.

Every weight was tuned, and every add-on kept or dropped, on the 2024 auction using only data from before it. Mega and mini auctions are never mixed, and the real 2026 results are read only by the comparison script.

Impact

For a franchise, AuctionIQ is a rehearsal room: who else needs this player, how high the room is likely to go, and where the bargains are. For fans and media, it explains an auction instead of just reporting it.

Challenges we ran into

Keeping the test honest: the 2026 results never feed back into the model.
Teaching agents to walk away: realistic bidding comes from needs and purse pressure, not just player value.
Playing a lot bid by bid in a 3D room smoothly on a phone.