Kynnys is Finnish for threshold

We measured the market before we priced it.

A dispatch desk that runs a car-transport week by itself: it measures what each dealer accepts and what each driver takes, then quotes the price with the best expected profit.

The final
EUR 17,463
Rank 22 of 58, 98.0% on time
62 unseen practice weeks
EUR 16.1k a week
62 of 62 ranked
Weeks under 90% on time
0 of 168
The Saturday system had 5 of 116
Test final 2
Rank 13 of 75
98.4% on time

Price tomography

Every request has one sharp threshold

We replayed the same weeks at ten price levels. A dealer does not drift from yes to no: each request flips at one sharp price. An online Bayesian tracker learns those thresholds during the week.

One request, ten prices

price / list price

Price ladder: the dealer says yes up to 1.3 times list and no from 1.4

Quote 1.20 x list and the dealer says yes.

One real request replayed at ten prices (price/list -> dealer said yes 1 / no 0). 97% of requests flip exactly once.

Where the threshold moves

Spread of the threshold, standard deviation in price / list

  • The day, in this week0.069
  • The lane, in this week0.056
  • Destination town, every week0.051
  • Origin town, in this week0.036
  • Destination town, in this week0.036
  • The dealer, in this week0.035
  • The lane, every week0.035
  • Noise of the single request0.110
fitted on past weekslearned live by the trackernoise

Fee tomography

Drivers have a threshold too

Then we ran the same test on the other side of the market: ten flat fee levels, 30,256 orders. Below the typical fee, jobs wait. Above it, extra money buys almost nothing.

The driver booking curve

booked within 60 minbooked before the deadline warning
Share of jobs booked rises with the fee and levels off above the typical fee

Fee experiment: 30,256 orders at ten flat fees. Fee / typical fee -> % booked within 60 min, % booked before the deadline warning.

The one rule

Quote the price that maximises P(dealer says yes) x profit.

For each delivery option, quote the price that maximises P(dealer says yes) x profit if we win; profit = price - expected fee - P(late) x cost of a late order - capacity charge + value of the dealer's trust.

Trained on GPUs

The learned part

Most of Kynnys is measurement and one rule. Two small models sit on top, each tested on weeks it never saw.

Late-risk neural net

in the bot

AUC, table → net

0.790.84

Log loss, table → net

0.1930.166

+2.5% margin on held-out weeks

Small neural net (8 nets, 2 layers of 16, 21 inputs) on top of the cost table, trained on 154,747 orders from 440 weeks on 8 RTX PRO 6000 GPUs, cross-validated by week.

Per-job fee model

next step

Value per job: best flat fee → per-job fee

EUR 38.98EUR 42.56

+9% offline, not yet in the bot

Per-job fee model (A100). Offline policy value on the randomised fee experiment.

One real practice week

Replay a week

One practice week (level 8, seed c3) played by the final system (commit 6b085575). One row per price request. Press play: about 20 seconds from Monday morning to Friday night.

Monday

07:00

Price requests

0

Won

0

Win rate of quotes

-

Declined

0

Timeline of the week: each price request is a dot, placed by time and by quoted price over list price
wonquoted, lostdeclinedheight = quoted price / list price

    Week result: EUR 17,154 margin, 435 orders, 96.1% on time.

    No step on faith

    Each step, measured

    Gross margin in thousand EUR over the same five level-8 practice weeks (seeds a1-a5).

    1. First working desk35.7k · 87.0%
    2. + quotes aim at a target win rate51.1k · 94.8%
    3. + fitted cost table and the one rule58.1k · 92.2%
    4. + price tracker, capacity charge, fee ramp63.0k · 94.6%
    5. + note reader, weekly calibration, guard69.7k · 95.0%
    6. + flat fee by job kind, awake hours, sized charges82.1k · 94.1%
    7. + late-risk neural net, cheap-market mode, trust value83.3k · 94.9%