DEV Community

Evgeny Shumilin
Evgeny Shumilin

Posted on Edited on

Econophysics for Multi-Asset Execution: Probabilistic State Modeling and Portfolio-Level Control

An experimental approach to multi-asset execution based on statistical mechanics, market microstructure and probabilistic modeling. A three-trade demonstration using live exchange data.

these are just random values
Research and Demonstration Disclaimer
This work is presented strictly for quantitative and academic research purposes. All models, metrics and execution data are generated in a demonstration environment designed to reproduce real exchange market conditions as closely as possible, including order-book dynamics, latency, liquidity constraints and transaction costs.
The results shown here come from simulated trading, not live-money execution. Nothing in this article constitutes investment advice, a commercial offer or a guarantee of future performance.

  1. The Core Idea: Modeling Markets Through Physics The basic idea behind this system is to treat financial markets as open, non-equilibrium statistical-mechanical systems. Rather than training neural networks or machine-learning models on historical price data, the system uses explicit mathematical formulas to estimate the current market state and calculate probabilities. There are no neural networks, no conventional forecasting algorithms and no predefined sets of conditional trading rules. The calculations are based on measurable market characteristics and their relationships. The model considers several aspects of market behavior: Liquidity distribution and order-book pressure. Order-flow dynamics and changes in trading intensity. Stochastic volatility and anomalous diffusion. Probability-density evolution. Execution costs, including transaction fees and slippage. Every execution decision can be traced back to measurable variables, such as probability-density drift, local diffusion characteristics and the friction associated with entering and exiting a position. The working assumption is that these underlying market characteristics remain measurable even when market narratives and sentiment change. Whether this provides meaningful resilience across different market regimes is something that still needs to be tested.
  2. Multi-Exchange Data Ingestion The current engine collects real-time market data from three exchanges: Binance, Bybit and OKX. One of the technical requirements is that the system cannot rely on a single fixed data structure for every instrument. Different exchanges and contracts have different specifications, so the ingestion layer resolves these differences dynamically. For each asset, the system accounts for: Tick size, quantity increments and minimum notional requirements. Order-book depth and liquidity distribution. Market-order price protection limits and changes in spreads. Exchange-specific maker and taker fees. Funding-rate schedules, including the eight-hour intervals used in the current configuration. The system also compares order-book density and trade frequency across all three exchanges. This makes it possible to distinguish broader liquidity imbalances from activity that appears on only one trading venue. The intention is to build a common market-state representation without losing the characteristics specific to each exchange or instrument.

  1. The Eight-Asset Universe Each of the eight monitored perpetual contracts has its own set of probability parameters. The model does not assume that all assets behave in the same way. I’ve assigned each asset a working physical classification based on its observed liquidity characteristics, price dynamics and execution requirements.

  1. The Mathematics of Probability The central part of the model is the evolution of probability density. For short-term price movements, we define the logarithmic displacement as:


The evolution of this variable is described using a state-dependent Fokker–Planck formulation:

The system continuously evaluates changes in this distribution. A trade can be triggered when the probability mass becomes sufficiently asymmetric beyond the model’s confidence bounds, in either direction: Long or Short.
Diffusive Noise vs. Ballistic Movement
Another important part of the model is distinguishing ordinary price fluctuations from more persistent directional movements.
Under standard Brownian diffusion, the expected displacement scales approximately with the square root of time.
A value around (H=0.5) corresponds to ordinary diffusive behavior, while (H>0.5) indicates superdiffusive behavior, where displacement grows faster than expected under standard Brownian scaling.
Thermodynamic Friction and Energy Dissipation
The model also uses a thermodynamic analogy to describe the cost of execution.
In this terminology, favorable price displacement represents captured energy, while trading fees and slippage represent energy dissipation. These are modeling terms, not measurements of literal physical energy.
This is only the starting point. Actual execution costs also depend on spreads and slippage, so exceeding the fee threshold alone does not guarantee a profitable trade.

  1. Portfolio-Level Control Although each asset has its own probability estimator, the portfolio operates under a common allocation mechanism. There is one important restriction: only one position can be open at any given time. Sequential Allocation The system rotates its active risk budget across the eight monitored assets. It doesn’t open several positions simultaneously. This avoids stacking exposure across correlated assets and reduces the possibility of multiple positions losing value together during a broad market movement. It does not eliminate market risk or protect an individual open position from a systemic shock. The point is to limit simultaneous exposure. Multiplicative Capital Management The portfolio controller maintains a target effective leverage of approximately 10.0x equity. The current configuration uses 50x contract leverage, with 20% of total equity allocated as initial margin. The idea is to maintain a consistent relationship between portfolio equity and market exposure rather than using a fixed position size. Low Exposure Time The portfolio spends most of its time outside the market. Instead of maintaining continuous exposure, the system observes market conditions and deploys capital when it identifies a sufficiently strong probability imbalance. Once the modeled price movement has been captured, it returns to a flat position. The intended exposure profile is roughly 90% or more of session time spent observing rather than trading. The demonstration below gives us an opportunity to see what this looks like in practice.
  2. The Demonstration: The accompanying video shows a complete demonstration session running. The system executed three trades across three different assets: SOL — Short ETH — Short ZEC— Long

All market data was collected in real time from Binance, Bybit and OKX, while execution took place in the Bybit demo environment.
Trade-by-Trade Results

Price displacement is shown as the absolute magnitude of the favorable move. For the short positions, this means the underlying market price moved down.
The ZECUSDT trade had a particularly interesting profile. Its impulse index was approximately 6.4 times higher than that of ETHUSDT, despite lasting only 8.33 seconds.
This is one of the differences the model is designed to identify: an asset can produce a much stronger directional movement over a very short period, while another may require a considerably longer holding time to achieve a smaller displacement.
Portfolio-Level Results
Looking at the entire session rather than individual positions:

The initial equity was 960.80 USDT. After the three trades, the reported account balance reached approximately 986 USDT. One detail worth mentioning is that the account’s total exposure time was only 8.43% of the session. For the remaining 91.57%, the portfolio was flat and observing the market.

  1. What the Demonstration Tells Us There are several things worth examining in these results. First, the model captured profitable movements in both directions. The SOL and ETH positions were short, while the ZEC position was long. This gives us an initial example of bidirectional execution using the same underlying probability framework. Second, the portfolio controller maintained its sequential allocation. There were no simultaneous positions, and exposure was limited to a small fraction of the total session time. The position-sizing mechanism also maintained approximately 10x effective equity exposure across all three trades. Under the stated 50x contract leverage, this corresponds to around 20% of equity being required as initial margin. The remaining balance was outside the initial-margin allocation, although it should not be considered risk-free collateral. But there is one problem that stands out much more than the others. The Main Bottleneck: Transaction Costs Out of the $60.23 in gross gains, $35.05 went toward taker fees. That’s 58.19%. The current fee assumption of 0.0605% per side means a round trip costs approximately 0.121% of notional. At 10x effective portfolio exposure, this is equivalent to roughly 1.21% of portfolio equity per completed round trip. For slower, more diffusive movements such as SOL and ETH, the margin between the captured price displacement and execution costs is relatively narrow. This creates a clear engineering priority: improve execution efficiency without compromising the underlying model.

  1. What Comes Next There are three areas I want to investigate further. Maker-Coupled Entry Execution The first is replacing market taker entries with passive limit orders. The current model uses a taker fee of 0.0605% per side. The proposed configuration would use a 0.0200% maker fee for entries while retaining taker execution for exits. Under those assumptions, the round-trip fee rate would fall from 0.1210% to 0.0805%, reducing total fee friction by approximately 33.5%. On the same three price movements, with the same position sizes and assuming every passive entry gets filled at the same price, the estimated results would change as follows:

This is a mathematical comparison, not an observed result. Passive limit orders may not fill, may be filled only partially, or may introduce additional waiting time and adverse selection. The actual outcome needs to be tested.
Impulse-Velocity Filtering
The second direction is to investigate whether capital allocation can be weighted toward assets and setups with higher impulse intensity.
The ZEC trade is an interesting example. Its impulse index was significantly higher than the other two trades, and its net contribution was also larger.
I want to investigate whether emphasizing these short-duration directional impulses can improve the amount of net movement captured per unit of market exposure time.
Of course, higher impulse intensity can also come with greater tail risk, particularly in thinner markets. This needs to be accounted for in further tests.
Multi-Venue Smart Routing
The third direction is dynamic execution routing.
Since the system already collects data from Binance, Bybit and OKX, it makes sense to evaluate which exchange offers the lowest combined execution cost at the moment a position is triggered.
That means considering the instantaneous spread, applicable trading fees and expected slippage rather than simply sending every order to a predefined exchange.
The objective is to reduce friction while keeping the market-state calculations independent of the execution venue.
Final Thoughts
This demonstration is still an early experiment. Three trades over 48 minutes are nowhere near enough to establish statistical reliability or demonstrate consistent performance across different market regimes.
What interests me here is the interaction between the individual asset models and the portfolio controller. Each instrument has its own dynamics, but the capital allocation mechanism operates around the portfolio as a whole. The next step is to run much longer tests, examine different market conditions and find out whether the same behavior holds across a larger asset universe.
If our interests and expectations happen to align, you may be able to find my profile through Y Combinator’s Startup School Co-Founder Matching.
I’m also on web3.career.
And finally, there’s my email address ( shumilin.e@icloud.com ) — although I’d prefer to use it only as a last resort.
Thank you for taking the time to read this rather lengthy article.

Top comments (0)