Development Update - 3 Nov - 30 Nov 2022

This month, the engineering team has had its focus of attention predominantly on improving the existing agent based model (ABM) that we’ve been developing over the last couple of months to conduct economic simulations.

Much of this work has been carried out in conjunction with CADLabs, a specialist firm who design and validate complex Web3 systems. 

Simulation

Agent decision rule performance optimisation

‍

Our existing ABM used a simple implementation of decision rules that helped us to run our simulation but didn’t provide enough depth when it came to how agents took decisions in real life. 

This simple implementation involved using constant probability values. In addition to this the existing decision rules were not the most performant.

Hence our team working on profiling the simulation to improve the performance of the Decision rules, worked on designing and implementing a set of default and agent specific decision rules i.e different decision rules for taker and makers. 

This has allowed us to add more layers to our agent decision making and to make relatively realistic decisions. 

Without prematurely optimising, we lifted calculations applied to individual agents to being calculated at each timestep for all agents, for example:

• A 30 day moving average of ETH price to determine market sentiment of agents can be calculated at a global level
• For some calculations where only one component of the calculation is for example stochastic for each agent, that individual component can be calculated at agent level with the rest of the calculation performed at a global level
• Move the relevant decision factors at protocol level (e.g. protocol health, size) outside of each agent 

‍

Synthetic Options price generation 

‍

It’s theoretically possible to replicate the payoff function of Bumper using put options and other financial derivatives. 

Deribit (European put options) has been the focus of analysis until this point. Hegic also offers a similar strategy, but is less established. 

Additionally, during early competitor analysis, it was hypothesised that Bumper payoff could be replicated using perps - it is not clear exactly how the pricing (accounting for risk too) can be equated between Bumper and some of the more complex derivatives such that we can integrate these findings into the model for agent decision making (based on costs for takers and returns for makers) between Bumper and some other protocol.

The first step was to perform some exploratory data analysis (EDA) of Deribit and Hegic Options prices, possibly creating a volatility surface, with the objective of being able to intuitively understand and be able to explain how options pricing on these platforms differs from Black-Scholes.

Currently, the agent decision rules for both takers and makers makes a comparison to Black-Scholes.

In the process of EDA, we also aimed to determine:

• The total value locked (TVL) of these platforms, segmented by both Takers and Makers when possible.
• The total daily trading volume of these platforms, and also segmented by both takers and makers when possible.

The above were used in parameterisation of agent demand function.

As a result we performed an analysis, with relevant visualisations, and established a function for generating synthetic options prices based on ETH price and other relevant inputs at an approximately 25 minute resolution.

Designed the parameterisation process for premium components.

The components that contribute to the protocol premium serve two purposes:

• To ensure competitive pricing of premium
• To define a mechanism for ensuring protocol stability through market incentives.

The protocol premium thus needs to be conservatively parameterised for competitive pricing + protocol stability.

The scope of this work was to determine the best series of steps to design and test the parameterisation of the protocol premium.

e.g. it was suggested that some components can be tuned by selectively configuring the model in different ways, for example setting a static ETH price, where under these conditions we’d expect a certain outcome.

The key goal of these tests was to determine whether a premium can be found that offers both competitive pricing in normal market conditions, and enabled the protocol to respond to instability by incentivising an appropriate market response.

Other High priority updates:

• Update to RiskWeightedAverageFloorPrice 
• Review and refactorisation of protocol parameters
• Create steady demand agent behaviour which resulted in a linearly increasing number of positions.
• Full mode simulation profiling
• Refactor open branches
• Sigma-time solving analysis

‍

Sim Testing

Our simulation team has been performing some preliminary tests on the Bumper protocol. 

These included a Sinusoidal price path frequency and amplitude sweep to construct an analysis of the protocol’s ability to remain solvent through a range of price volatility.

The results of this testing provided a number of useful insights, not least that the protocol seems resilient to volatility. 

Furthermore, Preliminary results of parameters used in tuning the bumper premium to a BSM benchmark show close tracking is possible provided these are dynamically updated. 

It was ascertained that certain risk factors and parameters used in calculating Bumper’s premium play a more significant role than others - this was in accordance with our initial intuition!

‍

Parameterisation and analyses:

• We focused on careful selection and parameterisation of agent decision factors
• Performed increased stress testing following initial parameterisation of the model, to test the bounds of stability and sanity check design of the protocol overall, and further iterate on parameterisation
• Continued audit of protocol rebalancing
• Continued sanity checks of protocol dynamics

‍

Determine which parameterisation processes and analyses need to be scaled up, requiring server resources:

• Currently, a standard simulation without detailed per-timestep agent state consumes approx. 5-10GB of RAM, which is manageable
• Larger simulations, and those where the simulation runs until completion over 365 timesteps at 1 day per timestep take significantly more RAM
• Simulations using full-mode 15m timesteps will either need to be scaled down to shorter time frames or be run on a server
• Simulations with more granular agent positions (i.e. smaller position size resulting in more agents) will either need to be scaled down to shorter time frames or be run on a server

‍

Up next:

The following are the tasks which are listed as up next in the technical plan:

• Analysis & parameterisation of protocol target yield - Medium priority
• Stable potential demand - High priority
• Per-timestamp agent joining data structure (for diagnostics) - High priority
• Parameterisation of sacrificial Makers and Takers - Medium priority
• Ability to introduce step change in ETH price path - Medium priority

‍

‍