The Lab · Research Proposal · Paper in progress

Agentic Trading Under Pre-Registration: Momentum, Sentiment and Volatility Targeting at Retail Scale

MTUTHUKO MNGOMEZULU · ISINKWAMTASE RESEARCH · KWATHEMA, SOUTH AFRICA · 2026 · DRAFT v0.3

Abstract. Most retail algorithmic traders lose money, and most published trading results are survivorship-biased marketing. This study asks whether an autonomous AI agent, governed by pre-registered promotion criteria, can manage a deliberately small real crypto account without falling into the failure modes that destroy retail accounts: overtrading, tight-stop bleed, euphoria entries and unsized volatility. We describe an agentic audition protocol in which candidate strategies are generated and evaluated by fleets of AI research agents, must survive fee-inclusive backtests across nine years and ten assets, walk-forward windows spanning five market regimes, and bootstrap Monte Carlo analysis — before touching real capital, which itself remains in a withdrawal-disabled exchange sub-account. Five audition rounds (200+ backtests, 21 strategy variants) are complete — including a forensic audit of a real high-frequency retail bot whose live month paid 14.6× more in fees than it earned — and no variant has yet earned promotion, which we argue is the protocol succeeding, not failing. Preliminary findings include a large positive effect of volatility-targeted position sizing (identical signals: −62% → +229%), a sentiment gate that neutralised the 2021 mania cold-start, the persistent failure of short-selling and tight-stop styles, and the result that trade frequency, not signal quality, is the dominant predictor of retail account ruin.

1 · Background and motivation

Three forces motivate this study. First, the documented failure rate of retail algorithmic and leveraged trading — regulators worldwide report the large majority of retail derivative accounts lose money. Second, the arrival of capable AI agents makes it cheap to generate plausible-looking strategies at scale, which makes disciplined rejection machinery more valuable, not less: the bottleneck is no longer ideas but honest evaluation. Third, almost all public evidence in this space is promotional; experiments that publish their losses are rare. This project runs the entire lifecycle — research, simulation, live trading — in public, from a township-founded lab, on an account small enough that its loss is survivable and its lessons are general.

2 · Research questions

RQ1. Can any rules-based strategy, net of real fees, beat buy-and-hold across full crypto market cycles at retail scale?
RQ2. How much of strategy performance is attributable to position sizing rather than entry/exit signals?
RQ3. Does aggregated crowd sentiment (Crypto Fear & Greed Index) carry usable signal beyond price-derived features?
RQ4. Do pre-registered promotion gates materially change which strategies reach live capital, compared with common practice (deploying the best backtest)?
RQ5. Can a high-frequency technical-ensemble bot — intraday charts, multi-indicator scoring, Bayesian trade gating — achieve positive expectancy net of fees at retail scale, and does its live behaviour match a faithful backtest of its rules?
RQ6. Does a coin's unit price or liquidity tier ("cheap, high-volume altcoins") offer a retail edge that the majors do not, under the same rules and the same gates?

3 · Hypotheses

H1. Time-series momentum with a long-term regime filter outperforms the market on risk-adjusted terms across regimes, but not on absolute terms in strong bull windows.
H2. Volatility-targeted sizing improves compound growth and drawdown for identical signals (supported in Round 4: −62% → +229%, max drawdown halved).
H3. An extreme-greed entry gate improves cold-start performance in mania regimes and never harms full-path performance (partially supported in Round 4).
H4. Intraday ensemble signals contain gross edge that survives fees once gated selectively enough (rejected in Round 5: at every gate level tested, gross expectancy was ≈ 0 and net expectancy was below the fee, in bull and bear regimes alike).
H5. A liquid sub-$1 altcoin universe offers higher trend-following returns than the majors for identical rules (rejected in Round 6: eleven pre-registered variants on eighteen coins, seven walk-forward windows — every variant failed the gate, and the two-coin incumbent beat ten of eleven challengers out of sample).

4 · Methodology

Agentic audition protocol. Each round, a fleet of AI research agents backtests candidate strategies in parallel (freqtrade engine, fee-inclusive), extracts structured metrics, and a deterministic script — not a model — applies the gate.
Pre-registration. Promotion criteria are committed to code before results exist: profitability in and out of sample, beating the market in a supermajority of walk-forward windows, minimum trade counts, statistical significance (p < 0.10), and Monte Carlo loss-probability bounds.
Walk-forward evaluation. Yearly and half-yearly windows spanning the 2018 collapse, 2019 recovery, COVID 2020, 2021 mania, 2022 cascade (Luna/FTX), the 2024 ETF era and the 2025–26 bear.
Bootstrap Monte Carlo. 10,000 resamples of each finalist's realised trade sequence, yielding outcome distributions, loss probabilities and drawdown percentiles rather than point estimates.
Phased promotion. Backtest → live simulation (weeks, on live prices) → live capital, with the account held in a withdrawal-disabled AI sub-account (OAuth-scoped) as a structural safety bound.
Radical transparency. Every daily decision and its reasoning is published; losing results are reported with the same prominence as winning ones.
Live-system forensics. Where a real trading system exists, its complete live trade ledger (every fill, fee and stop) is audited alongside a faithful re-implementation of its rules, so live behaviour and backtest behaviour can be compared directly — Round 5 audited a production bot's 708 live trades this way.

Adversarial review of the harness itself. From Round 6 the backtest harness is treated as a hypothesis too. Each round's code is attacked by three independent reviewers with distinct lenses — lookahead, execution realism, data and universe integrity — and the best out-of-sample result is re-implemented from scratch without the backtesting library and compared trade by trade. Round 6's reviewers found that the benchmark had been credited with a final trading day the strategies were denied (worth eight points out of sample) and that the two engines ran at different exposures; both were corrected before the verdict was read, and the verdict did not change. The reproduction matched the library to four decimal places on every trade.

5 · Data

Daily and intraday OHLCV for ten major crypto assets, Aug 2017 – present (Binance spot; Bybit derivatives for two-sided tests) — the full liquid history of the asset class. Crypto is 16 years old; no longer series exists.
Crypto Fear & Greed Index, Feb 2018 – present (3,100+ daily readings) as the crowd-sentiment proxy.
All macro/geopolitical shocks enter through prices; event-study extensions are future work.

6 · Preliminary findings (Rounds 1–6)

#	Finding	Evidence
F1	Fees and overtrading, not signal quality, destroy small accounts first.	194-trade strategy burned ~27% of account in fees; 1,800+ trade tight-stop style lost 96%.
F2	Short-selling failed in every configuration tested.	Three rounds; e.g. two-sided regime flip −89% vs long-only −7% on identical filters.
F3	Position sizing dominates signal choice.	Identical entries/exits: −62% naive vs +229% vol-targeted, drawdown 97% → 51%.
F4	Crowd sentiment helps exactly where theory predicts.	Extreme-greed gate turned the 2021 cold-start window from −83% to +69%; no full-path effect.
F5	Nothing tested beats holding across full cycles; edges are not yet statistically significant.	Market +1,168% over 9y; best variant +229%, p = 0.47; promotion denied by pre-registered gate.
F6	A live retail bot's fee bill can exceed its profit by an order of magnitude while looking "profitable" day to day.	Audited production bot: 708 real trades in one month, net +$0.80, fees $11.65 (14.6× net); long book +$5.00, short book −$4.20.
F7	Trade frequency, not signal quality, is the dominant predictor of ruin; selectivity only slows the bleed.	Faithful port of the bot's rules: −96% to −100% in and out of sample (6,500–20,000 trades); at maximum gating still −57% on 12 trades/day. Gross edge ≈ 0 at every gate level, both regimes.
F8	Unit price is not an edge. "Cheap, liquid" altcoins are a volatility tier, not an opportunity tier.	18 sub-$1 coins, 11 pre-registered variants, 7 walk-forward windows: the two-coin incumbent beat 10 of 11 out of sample; the regime rule that made +119% on BTC/ETH in-sample made −55% on the alt basket.
F9	Altcoin-basket returns cluster in one or two half-years; a rule positive in ≤ 4 of 7 windows is a start-date bet, not an edge.	Best breakout variant: +43% in-sample from two windows (+32%, +81%) with losses in the other five; the basket itself went +100% → −47% → −48% across three consecutive half-years.

7 · Planned work

Round 7: machine-learning candidates (FreqAI) under strictly walk-forward training, same gates; and a daily-timeframe descendant of the Round 5 ensemble (the signal stack re-tested at the frequency where fees stop dominating). Round 6 (the altcoin-universe test, RQ6/H5) is complete and reported above.
Live phase study: tracking error between simulation and live execution at minimum order sizes.
Write-up: full paper covering protocol, results and the sociology of "profit in every trade" expectations; preprint first, then peer review.

8 · Ethics and disclosure

The study trades only the author's own, deliberately small capital in a withdrawal-disabled sub-account. It is not investment advice, offers no managed product, and solicits no funds. Published results include losses. Nothing in this study constitutes financial services under South African law.

DRAFT v0.3 · 2026-09-22 · Live experiment and daily decision journal: mngomezulu.africa/lab/agent-trading