What is Aegis?

---

Aegis is rather simple; implementation of base technicals; MACD, BB, RSI to create signals ("Signals" are simply the conventional 'breakouts' or 'bullish' intervals of all TIs, aggregated). Clean signals are passed through VMD decomposition to strip microstructure noise, then fed into Gemma 4 31B (no specialised fine-tuning) to estimate profitability, OHLVC is aggregated and simplified to variables through a scoring system [high - probability of profit (POP): 0.5<, Low - POP: 0-0.5, max: 1), the estimate is continously refined via a bayesian updater, execution is routed through Alpaca API and only proceeds if P(Profit) > 60% and net yield clears transaction costs.

---

System Architecture

flowchart TD
    A[(NASDAQ\nTick Feed)]
    
    A --> B[OHLCV Aggregation\nOrder Book Depth]
    B --> C[Technical Indicators\nRSI · MACD · BB]
    C --> D[VMD De-noising\nSignal / Noise Separation]
    
    D --> E{Volatility > 3σ ?}
    E -- Yes --> HOLD
    E -- No --> F[BiGRU\nProfitability Prediction]
    
    F --> G[Bayesian Updater\nPosterior ∝ Likelihood × Prior]
    G --> H{P Profit > 60% ?}
    
    H -- No --> HOLD
    H -- Yes --> I{Net Yield\nAfter Costs > 0 ?}
    
    I -- No --> HOLD([Hold / No Edge])
    I -- Yes --> J([Execute\nLimit / Market Order])

Loading

---

Central Component

[STATUS: NON-VIABLE]

---

Raw tick data is decomposed via VMD minimisation to isolate tradeable signal components from market microstructure noise:

$$f(\alpha, \beta, \gamma) = \min_{k,,\beta,,\gamma} { | \delta(t) - \sum_{k}^{n} u_k(t) | + \alpha | \partial_t [ u_k(t),e^{-i\omega_k t} ] | }$$

Extracts discrete sub-signals $u_k$ from raw input $\delta(t)$. The penalty term $\alpha$ regularises the derivative of each frequency-shifted mode, suppressing high-frequency noise without distorting the underlying signal structure.

---

Post-BiGRU, the system applies standard Bayesian inference over a rolling state:

$$\text{Posterior}(\theta \mid y) ;\propto; P(y \mid \theta),P(\theta)$$

Each inference cycle maps the prior $P(\theta)$ against the likelihood of observed market data $P(y \mid \theta)$ to yield a continuous posterior distribution over signal validity. The posterior is recycled as the next prior, maintaining an adaptive belief state across time.

---

Two hard gates govern execution eligibility:

Gate	Condition	Description
Volatility Filter	$\sigma_t \leq 3\sigma$	Suspends execution during fat-tail regimes; no edge assumed beyond 3 standard deviations
Execution Gate	$\psi^* \xi > 0.6$	Bayesian posterior $P(\text{Profit})$ must exceed 60% before order routing proceeds

---

Primary bottleneck: Pronounced performance decay during structural regime shifts, Bayesian prior state exhibits adaptability drag under non-stationary volatility, causing posterior miscalibration. Regime change analysis ongoing.

---

Sources

1. Cont, R. & Larrard, A. — Dynamics of Order Positions and Related Queues in a Limit Order Book https://www.researchgate.net/publication/277023747_Dynamics_of_Order_Positions_and_Related_Queues_in_a_Limit_Order_Book

2. Adams, R. P. & MacKay, D. J. C. — Bayesian Online Changepoint Detection https://arxiv.org/abs/0710.3742

3. Cho, K. et al. — Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation https://arxiv.org/abs/1406.1078

4. Linux Kernel Documentation — AF_XDP https://docs.kernel.org/networking/af_xdp.html

5. DPDK — Poll Mode Driver (v24.03) https://doc.dpdk.org/guides-24.03/prog_guide/poll_mode_drv.html

6. DPDK — OPDL Event Device (v24.07) https://doc.dpdk.org/guides-24.07/eventdevs/opdl.html

7. DPDK — ENA NIC Driver (v25.03) https://doc.dpdk.org/guides-25.03/nics/ena.html

8. DPDK — ICE NIC Driver (v25.11) https://doc.dpdk.org/guides-25.11/nics/ice.html

9. DPDK Power Docs — ENA NIC https://dpdk-power-docs.readthedocs.io/en/latest/nics/ena.html

10. NVIDIA — TensorRT Developer Guide https://docs.nvidia.com/deeplearning/tensorrt/developer-guide/

11. AMD/Solarflare — Enterprise Onload User Guide (SF-104474-CD-34) https://www.amd.com/content/dam/amd/en/support/downloads/solarflare/onload/enterprise-onload/SF-104474-CD-34_Onload_User_Guide.pdf

12. Lahmiri, S. — A Variational Mode Decomposition Approach for Analysis and Forecasting of Economic and Financial Time Series https://www.researchgate.net/publication/295394724_A_Variational_Mode_Decompoisition_Approach_for_Analysis_and_Forecasting_of_Economic_and_Financial_Time_Series

13. Zhang, Z., Zohren, S. & Roberts, S. — DeepLOB: Deep Convolutional Neural Networks for Limit Order Books https://ieeexplore.ieee.org/document/8673598

14. Jin, M. et al. — Time-LLM: Time Series Forecasting by Reprogramming Large Language Models https://arxiv.org/abs/2310.07820

15. Lopez de Prado, M. — Advances in Financial Machine Learning https://www.wiley.com/en-us/Advances+in+Financial+Machine+Learning-p-9781119482086

16. Fowler, M. — The LMAX Architecture https://martinfowler.com/articles/lmax.html

17. Drepper, U. — What Every Programmer Should Know About Memory https://people.freebsd.org/~lstewart/articles/cpumemory.pdf

18. Belay, A. et al. — IX: A Protected Dataplane Operating System for High Throughput and Low Latency https://www.usenix.org/system/files/conference/osdi14/osdi14-paper-belay.pdf