predicting startup growth with hierarchical gaussian process

[hyunjimoon]

applying https://mc-stan.org/users/documentation/case-studies/gaussian-process.html, a casestudy which predicts failure counts based on "highly unbalanced per category with lots of missing data" and "hierarchical structure of a system, such as shared engine types", we predict startup growth. startup growth is also highly unbalanced (more relevant data in fast clockspeed industry (IT) than slow clockspeed (biotech, semiconductor manufacturing), lots of missing data, hierarchical structure (biotech (gene therapeutics(hardware), bioinformatics (software) and semiconductor (AI accelerator chips (hardware), Electronic Design Automation (software)).

• hierarchical: EDA tools are crucial for designing AI accelerators and other semiconductor devices, bioinformatics is critical for the research, development, and application of gene therapies, offering the computational power to analyze and interpret biological data.
• Length Scales representing the different impacts of technological and operational actions on startup growth.
• Scale Variance showing the variability in growth due to strategic goals and customer development.
• Periodicity reflecting the recurring events like partnership formations and funding rounds.
• Covariance Between Actions indicating the interdependencies of different AGCT actions.
• Non-Stationarity and Multimodal Characteristics to depict the evolving influence of actions and the various growth phases startups experience.
• Integration of Discrete Events highlighting significant milestones like regulatory approvals.
• Latent Variables suggesting the hidden factors impacting growth.

[hyunjimoon]

⚽️ SOCER matters!

Parameter	Hypothesis on Time to Change Market over Change product (T2CMP) w.r.t. Parameter	Examples in EV	Examples in Autonomous Vehicles
S: Stability of market	The more dynamic market (low S) is, the less it updates its belief about market acceptance, hence longer it takes to pivot the market.	An EV startup operating in a highly dynamic market with rapidly evolving customer preferences and competitor moves may take longer to pivot, as the changing conditions make it harder to draw firm conclusions about market fit.	An AV startup in a market with frequent regulatory shifts and public perception changes may require more time to pivot, as the unstable environment complicates assessing the true state of market acceptance.
O: Optimism (gap of belief and real on market acceptance)	The more optimistic a startup is about market acceptance, the less it updates its belief about market, hence longer it takes to pivot the market.	An EV startup that is highly optimistic about the market potential for a new battery technology may continue pursuing it longer before considering a pivot, as the positive outlook reduces the perceived need to update beliefs based on market feedback.	An AV startup with a very optimistic view of the readiness of its technology for a specific use case may persist longer before pivoting, as the belief gap makes it less responsive to signals challenging its assumptions.
C: Confidence on market acceptance	The more confident a startup is about the market acceptance (compared to production feasibility), the less it updates its belief about market, hence longer it takes to pivot the market.	An EV startup highly confident in the demand for a specific battery chemistry may stick with it longer before considering pivoting to a different technology or market segment.	An AV startup with strong conviction about the market acceptance of a particular use case, like autonomous shuttles, may persist longer before entertaining a pivot to another application.
E: Experiment opportunity	The more experiment opportunity a startup has, the more scientific experiments can be done (e.g., in parallel), hence it can pivot to a more precise market, if needed.	An EV startup with ample resources to conduct multiple market validation tests simultaneously can more quickly zero in on the most promising segments and pivot accordingly.	A well-funded AV startup that can afford to test various use cases and market segments in parallel will be able to more swiftly identify the optimal pivot direction based on comprehensive data.
R: Risk endurance	The more risk endurance a startup has, the more scientific experiments can be done (e.g., in parallel), hence it can pivot to a more precise market, if needed.	An EV startup with high risk tolerance can conduct bolder experiments testing novel battery chemistries and market applications, enabling a more informed and targeted pivot if required.	An AV startup with significant risk appetite may pursue more ambitious testing across a wider range of operating environments and use cases, allowing it to pivot more effectively to the best-fit market based on the data collected.

To shape this into transportation class project, I constructed tree in transportation subdomain. Matt Puchalski introduced me to Andrew, in the hope that he can share insights on whether there might be some transferrable components in startup operations across, say EV to AV industry. Once this is nailed we can extend this by going one level up in the tree (between pure software and software+hardware)

---

In transportation class, I learned zero-shot transfer learning can be applied in transportation to manage traffic signals across varying speed limits without training a separate model for each scenario. It focuses on employing a small number of models trained on specific speed limits to generalize across a range of speed limits using zero-shot transfer, where a model adapts to manage slightly different speed scenarios without direct retraining. Transferring this to early stage strartup experiments , we can think of sub industries as different speed limits (e.g. 5m/s for ride hailing, 20m/s for AV, 24m/s for `EV'.

For systematic divergence metric, I built a tree considering industry affinity within transportation for startup experiment sequencing..

The mobility startup industry encompasses a wide range of subdomains, each focusing on different aspects of transportation and mobility. Some of the key subdomains include:

1. Software (service branch from figure)

• ride-hailing and sharing: Companies like Uber, Lyft, and Didi Chuxing that connect passengers with drivers for on-demand transportation.
• Micromobility: Startups offering short-distance transportation solutions, such as bike-sharing (e.g., Mobike, Ofo), e-scooter sharing (e.g., Bird, Lime), and e-moped sharing (e.g., Revel).
• Car-sharing: Platforms that allow users to rent cars for short periods, such as Zipcar, Car2Go, and Getaround.

2. Involves hardware (manufacturing branch from figure)

• Autonomous vehicles: Companies developing self-driving technology for cars, trucks, and other vehicles (e.g., Waymo, Cruise, Aurora).
• Electric vehicles: Startups focused on designing, manufacturing, and selling electric cars, bikes, and scooters (e.g., Tesla, NIO, Rivian).

Remaining question: "WHICH DATA TO COLLECT"?

[hyunjimoon]

Cathy, Angie

• learn contextual mdp (for different industry)
• validate either input (is arrival poisson) or output (out of sample prediction accuracy)
• jh (purely based on simulated data so methodology contribution) and hj (given optimal action in different context, what is latent covariance matrix between different environment?)'s approach is opposite
• jh: given context (task), output policy's covariance
• hj: given policy's implementation, infer context's covariance matrix

[hyunjimoon]

junghoon, angie agenda

• pomdp modeling of early stage startup experiment / market-product pivot

• transfer learning between market-product pivot btw AV, EV industry table in predicting startup growth with hierarchical gaussian process #200 (comment)