DataQuant Consulting @DataQuantAI

Collections Scoring with Gradient Boosting An AR team can chase maybe 20% of overdue invoices in a given week. The question is which 20%. Sort by days-overdue and you waste the week on invoices that were going to self-cure. Sort by amount and you miss the small accounts that genuinely won't pay. A gradient-boosted model on payment history — days since last payment, payment velocity, dispute flags, seasonality — ranks invoices by recovery risk. Plot the cumulative gains and the value of ranking becomes obvious. Illustrative run: working the top 20% by model score captures ~58% of at-risk value, versus 20% if you chase at random. Same headcount. Same hours. Roughly 3× the recovered cash.

If your price elasticity estimate is around -0.4, check for confounders before you celebrate the inelastic demand. Naive regression biases elasticity toward zero — the real number is often 2-3× larger.

Price Elasticity with Double ML If you regress quantity on price to get elasticity, your number is almost certainly too small — biased toward zero. Why: price moves with the things that also move demand. Promos, season, competitor actions. The naive regression can't tell the price effect from the confounder effect, so it splits the difference. Double ML fixes this by orthogonalization: model demand from the controls, model price from the controls, then regress the residuals on each other. What's left is the part of price variation that isn't explained by confounders — the causal effect. Illustrative run: naive OLS says -0.42. Double ML recovers -1.29, right next to the true -1.35. That 3× gap is the difference between leaving margin on the table and pricing to the real curve.

Tail Spend, The Hidden Drain 20% of suppliers consume 80% of admin cost. Mostly invisibly. By spend value: strategic suppliers dominate. By admin cost: tail suppliers dominate. Both inversions are real and simultaneous. Auto-classification + catalog buying + PO-free thresholds: ↓ 40-60% tail processing cost.

The Spend Cube Category × supplier × geography × time. This is the data substrate every other procurement decision sits on. ~70% of mid-market companies don't have a clean one. Which means every procurement initiative starts with a 6-month data project — usually unbudgeted.

Supplier Risk Scoring Four signals. One number per supplier. Refreshed weekly. Financial — z-score, credit trend, payment behavior. Operational — OTIF, quality, capacity. ESG — emissions, labor, governance flags. Concentration — % of your spend × % of their revenue. Early warning 60-90 days before the disruption phone call.

Dynamic discounting is the highest-return idle-cash deployment most treasuries don't run. The math is trivial. The blocker is that AP is measured on payment timing, not on net cash return.

Three-Way Match, Automated PO. GRN. Invoice. Match → post and pay. Exception → route by category. 75% straight match in a typical mature rollout. 25% goes to exception queues — pricing variance, qty mismatch, missing GRN, tax error — each routed to the right team in seconds. Exception handling time ↓ 60%.

The hardest part of an AI rollout in finance isn't the model. It's the controllers learning to trust output they didn't generate. That trust takes one bad month to lose.

Dynamic Discounting Agent When paying early beats holding the cash. Discount rate annualizes above 12% APR → PAY EARLY. Marginal discount, strategic supplier → RELATIONSHIP PAY. Weak discount, non-critical supplier → EXTEND. Rule-based dynamic discounting: 150-300 bps annualized return on idle cash. Better than most short-term treasury options.

Churn doesn't surprise companies with good data. It surprises companies with disconnected data. The early warning is in the join.

Pricing models pay back in compounding margin. Q1 looks like nothing. Q4 looks like a different business. Most pilots get killed in Q2.

What if you could take three completely different model families… and distill them into one tiny model? 🤯 📜 Paper: arxiv.org/pdf/2605.21699 MOPD (Multi-Teacher On-Policy Distillation) has become a standard procedure in post-training. We already distill multiple specialized variants of the same model into a single set of weights. But what if we could go further - and distill models from entirely different families? Turns out, it is possible. Today we’re releasing a paper on cross-tokenizer distillation - our first steps in this exciting direction. 📄 We distilled Qwen3-4B, Phi-4-Mini, and Llama-3B into Llama-3.2-1B. MMLU jumped from 32.05 → 46.32 when using multiple teachers. 📈 The team is now working on Nemo-RL integration so the community can try this method in their own settings. Plus, we are scaling experiments up. 🚀

Budget-aware Agents (BAGEN) study the failure modes in budget estimation: 1. Strong agents are not strong budget estimators. 2. Frontier models are often overoptimistic. 3. Budget awareness is actionable and trainable. SFT plus RL strengthens early stop and alert behavior, saving 28-64 percent of tokens on failed trajectories. 4. Upper and lower bound calibration remains hard. ragen-ai.github.io/bagen/

Zihan "Zenus" Wang @wzenus

6 days ago

🧵 Claude-Opus-4.8 takes you too much tokens - but is this issue general across agents? Do agents know how much they'll spend? Introducing Budget-Aware Agents (BAGEN): We study budget awareness across 4 envs & 5 frontier agents, and find structured failures in most of them.

22 43 336 565K 153

Le labo de recherche General Robotics Lab de l'université Duke 🇺🇸 présente « Argus », un robot sphérique novateur conçu pour interagir avec son environnement de manière inédite (omnidirectionnel, et bâti sur le concept de « symétrie dynamique »). generalroboticslab.com/Argus

The Pricing Causal Model in Production What the deployed pricing model actually does each week: Sun — Feature refresh. POS, weather, competitor scrapes. Mon — Causal recompute. Elasticity by segment × SKU. Tue — Recommendations with confidence intervals. Wed — Pricing committee reviews top 50 SKUs. Thu — Approved moves push to ERP. Margin lift compounds over quarters as the model learns and the committee trusts. +3 to 7 margin points annualized by Q4. Not because the model got smarter. Because the loop got tighter.

The Variance Attribution Agent Before: Mon 9am — close locked, no commentary yet. Five days of controllers chasing data from eight systems. Board pack distributed Mon+1. After: Mon 9am — variance bridge + draft narrative on CFO's screen. Mon 11am — controller edits the why, not the what. Mon 3pm — board commentary done. Same data. Different posture. The agent doesn't replace controllers. It moves their work from assembly to analysis.

The procurement function reports savings to the board. The CFO never sees those savings land. Both are doing their job. The system between them is the problem.

Supplier risk scoring without concentration as a dimension is incomplete. It's the one that explains the most disruptions — and the one most score cards skip.

Negotiated ≠ Realized Where promised savings disappear before they hit the P&L. 100% negotiated → 85% after scope creep → 72% after volume shortfall → 60% after maverick spend → 55% after contract expiry gap → 50% realized. Half of every negotiated dollar never shows up. Realization tracking recovers most of it.