Pricing · Economics · Data Science
From wholesale market signals to governed pricing decisions.
I designed and built an end-to-end Pricing Intelligence system from nine years of Chinese wholesale vegetable prices. It measures market position, forms price expectations, screens cost risk, and evaluates sourcing scenarios—while refusing to release claims that fail their evidence gate.
Data-free public view
Historical cutoff: 2022-06-22
No automated pricing
Frozen out-of-sample evidence
Performance is shown against the decision rule.
These visuals use model-card aggregates already disclosed in the repository. No city-, product-, or date-level price observations are included in the public package.
P2 · Forecast
Forecast release by horizon
Each bar contains ten frozen product–horizon decisions. The 14/28-day horizons are the formal MVP decision focus; 7 days remains a stress test.
7d
82
14d
8
28d
62
Baseline fallback
Point model
Minimum + interval
Target + interval
Reading: 7 days averaged −4.24% relative WAPE improvement and fell back for 8/10 products. At 14 and 28 days, 9/10 products released model points, supporting the 14/28-day product focus without hiding the failed short-horizon test.
P3 · Alert
Alert stability: validation → final
The action threshold was frozen on validation. Final test measures drift without selecting a new threshold.
Recall 42.15 → 37.69%
Average precision 21.28 → 22.04%
FPR 9.96 → 9.99%
Reading: ranking quality holds, recall softens, and the false-positive constraint remains reproducible.
P5 · Methodological win
How controls stopped 99.9% of candidates from becoming claims
The method wins when the product refuses a dense, non-replicating network. Two distinct checks prevent shared movement and multiple testing from masquerading as city-to-city information.
01 · Confound check
Remove shared market movement
Before common-factor control
1,481
contemporaneous FDR relationships
−47.6%
after seasonality + leave-one-out common factor
After control
776
relationships still descriptive, not directional
02 · Direction claim gate
Require incremental, out-of-sample value
- 2,736predeclared geographic candidates
- 991estimable pairs
- 174raw train-significant
- 51after product BH-FDR
- 15validation + stability frozen
- 4positive final-test gains
26.7% positive vs 70% gate
−1.39% median RMSE uplift vs +1% gate
Governance decisionWithhold the network
0directional edges released
common_shock_only · common movement and city exposure remain available for human reviewWhy it matters: without common-shock control and product-level multiple-testing discipline, more than a thousand apparent relationships could have looked like a successful network. The sealed final test showed they were not shippable directional evidence.
Visual evidence only · Frozen historical evaluation · No row-level price data
Decision chain
Every stage earns a product state.
The project starts with the user decision and available information set, then chooses the simplest method that can clear a predeclared evidence bar.
Data release
Can these prices be compared?
Versioned market mapping, quality flags, coverage tiers, and two-level aggregation create auditable market and city facts.
Historical
Where is price relative to peers and season?
Trends, rankings, dispersion, seasonality, and source reliability support historical market benchmarking.
Partial release
Does forecasting beat a simple rule?
Rolling-origin 7/14/28-day tests improved overall WAPE by 6.30%. The product focuses on 14/28 days; 7 days remains a disclosed stress test with frozen fallbacks.
Alert release
Is a price increase worth reviewing?
A validation-frozen threshold achieved 37.69% recall at 9.99% FPR. Alerts open a review queue; they never change prices.
Scenario release
Does a low quote survive economic friction?
Risk-adjusted landed cost and 36 transport, loss, and risk scenarios identify cities worth requesting a quote from.
Network no-go
Can an apparent network survive confound and replication tests?
Common-factor control removed 47.60% of contemporaneous FDR relationships; sealed final testing then rejected the directional claim. Withholding the network is the result.
Evidence before sophistication
The strongest result is knowing when not to release.
Statistical significance, an attractive chart, or average improvement does not automatically become a product feature. Baselines, false-action cost, uncertainty, and replication determine the release state.
Confound control1,481 → 776
47.6% of apparent synchrony removed
Replication gate15 → 4
only 26.7% showed positive final gain
0 shippedNetwork withheld; common-shock evidence retained.
Last price, weekly pattern, seasonality, and target-history baselines remain visible when the advanced model loses.
P3 freezes its action threshold under a 10% false-positive-rate budget and discloses precision of 21.31%.
P4 separates the cheapest modeled option from a source that remains stable across 36 parameter combinations.
P5 uses common-shock control, product-level BH-FDR and sealed final testing to stop an attractive but non-replicating network from shipping.
Economics in the implementation
Economic reasoning changes the product behavior.
This is not an economics label added after modeling. The mechanisms determine aggregation, baselines, constraints, uncertainty, and which actions the interface permits.
Price dispersion & measurement
Market composition and sparse observations can imitate economic differences, so quality and coverage enter before comparison.
Expectations & uncertainty
Forecasts are evaluated against what was knowable at the time; points, intervals, and baselines receive separate release decisions.
Transaction costs & risk
A lower observed price is actionable only after distance, transport, loss, reliability, and uncertainty are applied consistently.
Common shocks & identification
National supply and seasonal movement compete with propagation stories; without causal identification, the system makes no causal claim.
Pricing workflow interface
Complementary to a Pricing Engine—not a competing engine.
This project supplies governed evidence upstream. A production Pricing Engine still owns the executable decision.
Pricing Intelligence supplies
- market benchmark and reliability
- model, baseline, or no-go status
- price-risk review trigger
- risk-adjusted cost scenario
- common-shock exposure
Pricing Engine combines
- current cost and inventory
- customer and contract terms
- demand response and margin floor
- commercial rules and approval
Then—and only then—produce an executable price.
What I built
Research discipline translated into product engineering.
One repository connects data lineage, temporal evaluation, model governance, decision interfaces, six web routes, career-ready documentation, and code-only continuous integration.
Data & analytics
Python · pandas · NumPy · PyArrow · SciPy · Parquet · versioned marts
Model evaluation
Rolling-origin backtests · conformal intervals · logistic risk ranking · BH-FDR · block bootstrap · sealed final tests
Product & quality
React · Vinext · TypeScript · Recharts · responsive UI · 178 tests · lint · GitHub Actions
Production boundary
What this evidence does—and does not—support.
Supported
- Historical market benchmarking
- Qualified out-of-sample forecast slices
- Manual risk-review prioritization
- Parameter-dependent sourcing inquiry
- Common-shock exposure analysis
Not supported
- Current or real-time recommendations
- Demand elasticity or willingness to pay
- Optimal retail price or realized savings
- Supplier capacity or executable orders
- Causal price propagation
Audit the work
Follow every public claim back to its evidence.
中文摘要
这是 Pricing Engine 的上游情报层，而不是另一套自动定价引擎。
项目把批发市场数据转化为历史基准、价格预期、风险复核、采购成本情景和共同冲击证据。最终价格仍由现有 Pricing Engine 结合库存、客户、合同、利润和审批规则决定。数据截至 2022-06-22，所有公开页面均不包含原始或派生数据。
Portfolio takeaway