Trading performance & risk audit

RESEARCH SNAPSHOT · No strategy activation

Trading audit — decision brief

September 10, 2026, 3:35–3:43 PM EDT (America/New_York). This is a fixed research snapshot, not a live quote screen. Refreshing this report does not update the audit. Live strategies and orders were not modified by this analysis.

The available data does not establish a reliably profitable replacement strategy. The strongest actionable findings are concentrated ETC exposure, losses larger than average wins, paper profit giveback, and incomplete cost-basis/history reconciliation. Paper results cover only one partial week.

Current strategy results

Real, demo and local paper money are kept separate. Combined live is a provisional sum of matched closures, dominated by Robinhood. Drawdown here is peak-to-trough cumulative CLOSED-trade P/L starting at zero, not total account or intratrade drawdown. Full account drawdown needs historical equity and cash-flow records. Zero closures means insufficient evidence, not a proven zero-loss strategy.

All accounts and combined portfolio

Cash definitions differ: Robinhood exposes crypto buying power, used as available cash; eToro credit is cash but no distinct buying-power field was returned. Coinbase includes USD/USDC and a wallet-only MINI position; MINI value is not necessarily executable on Advanced Trade. Paper cash is reconstructed from a $10,000 bankroll. Small rounding differences between position sums and broker equity are expected.

Today P/L and Week P/L are explicitly Unknown for every account in account-summary.csv. The existing live account-change display is a change in valuation from its stored baseline; deposits and withdrawals can affect it. It is not interchangeable with realized trading profit. Do not sum currencies or paper balances into real capital.

Robinhood results

All 266 returned crypto orders were paginated: 264 filled, 2 failed; 156 buy orders and 108 sell orders. Executions were aggregated per order, then matched FIFO to prior buys; repeated bot log lines were not counted again. Matched net is -$18,048.29, with a 64.81% win rate but average win $430.94 versus average loss $1,268.78. This is a provisional reconstruction, not certified lifetime or tax P/L: two sales lack complete earlier acquisition lots, and reconstructed BTC/SOL balances do not reconcile to current holdings. The integration does not expose Robinhood equities.

The old DOGE tracker reports WAITING even though DOGE was sold and proceeds rotated into ETC. That describes a legacy symbol-specific tracker, not the whole account. The audit holdings table uses broker holdings and clearly shows ETC. Existing ETC protection checks approximately each minute: -4% stop, +5% activation, then 3% peak drop. Those polling rules do not guarantee an execution price.

eToro results

Real: one returned completed position, net -$0.01; two BTC positions open. Demo: 182 broker-reported closures, net -$255.88. The demo strategy targets +0.25% versus a -1% stop: gross reward is one quarter of planned loss before fees/spread. Broker netProfit was used instead of adding duplicate local estimated P/L. Both APIs were exhausted within the requested 364-day window; earlier account lifetime was not established. Current positions report broker-native stop/target disabled; bot-owned positions rely on scheduled app checks. The manual real lot is outside automatic bot ownership.

Other account results

Alpaca: one QQQ acquisition, no completed round trip. Equity $992.51, cash $500, reported open loss about $7.48. Its weekly SPY/QQQ selection uses long-term trend, momentum and volatility with a $500 allocation cap; the emergency monitor runs during configured weekday hours. Coinbase: five Advanced Trade fills, one matched closure -$0.02. Existing legacy/wallet holdings predate this returned trading history; holdings cost basis from the portfolio API is available but does not reconstruct all historical sales. Coinbase cash includes $100 USD and about $0.0174 USDC: the bot checks spendable USDC against its reserve, explaining why cash can exist while it refuses a new USDC purchase. No funds were moved.

DEX paper: 55 exits, -$927.84. Coinbase mover paper: 17 exits, -$673.78. Canonical PAPER SELL events were counted once. Some DEX symbols collide across tokens/overlapping positions: exit P/L is retained but ambiguous entry/holding details are marked unknown. DEX simulates 1% entry and exit slippage; mover paper does not provide a complete fee model. Neither is equivalent to achievable live fills.

History coverage and data quality

Files include per-trade entry/exit/quantity, size, duration, P/L, percentage, returned fees and exit reason. Unknown fees, spreads, entry signals, stop activity and historical slippage cannot be recreated from a final execution alone. Robinhood effective prices already embed spread; zero separate fee is not zero trading friction. Transfers/deposits/withdrawals are not treated as trade P/L. A complete lifetime reconciliation still requires missing transaction/transfer and opening-lot records; additional accounts outside the configured credentials are not established.

Winning trade patterns / losing trade patterns

Highest net asset groups (descriptive; unequal samples and history windows).

Lowest net asset groups (descriptive; unequal samples and history windows).

Robinhood demonstrates why win rate is insufficient: its average losing trade is almost three times the average winner. Both paper scanners buy already-rising assets, exposing entries to reversals after momentum spikes. DEX requires liquidity/volume/transaction filters plus a second confirmation; mover requires roughly 5–10% 15-minute momentum plus confirmation. Entry-indicator snapshots are not retained for every closed trade, so causal claims about a particular threshold, market direction or volatility regime are not supportable. trade-patterns.csv shows descriptive weekday, holding-time and exit-reason buckets; these are not independent validation or a recommended trading schedule.

Profitable trades that turned into losses

47 paper trades have a recorded positive peak above their final result; 24 ultimately closed negative. Total observed giveback $4,939.92; average $105.10. This measures the difference from sampled peaks, not money that could necessarily have been captured. Historical maximum profit is unknown for live trades. Summing perfect exits at all peaks would introduce hindsight bias.

profit-giveback.csv contains every detected example, entry, duration and data-quality note. Proposed exit-only simulations below quantify conditional changes without assuming an achievable sale at the observed peak.

Current open positions

All are REVIEW at this snapshot: the audit does not have enough continuous lifecycle and execution evidence to prescribe an immediate exit or partial sale. open-positions.csv includes quantity, entry/current prices, percent P/L, sampled peak when available, existing stop/target/trailing status and holding time when known. These analytical labels did not create orders. An old snapshot should not be used as a current execution instruction.

Duplicate / correlated exposure

ETC is the dominant risk, far larger than the combined small BTC holdings across Robinhood, Coinbase and eToro. Multiple crypto assets share broad crypto-market exposure; a statistical correlation matrix cannot be established from these sparse, nonaligned records. QQQ adds equity-market exposure, not guaranteed diversification. Cash is excluded from invested exposure; eToro demo and paper holdings are excluded from live exposure.

Backtest results — targets, protection and loss controls

84 exit plans × two OHLC order assumptions × three incremental per-side costs = 504 scenarios. Fifteen of 17 Coinbase mover paper trades have usable interior one-minute candles; two very short trades are excluded. Entry and exit boundary minutes are excluded to prevent using prices before entry or after exit. DEX and broker histories lack comparable paths, so their exit-target backtests are unavailable. Recorded baseline on the same 15 trades: -$431.63, win 40%, PF 0.662, closed-trade drawdown $574.93.

Tested dollar targets $3/$5/$7.50/$10/$15/$20 against dollar stops $5/$10/$15/$20, and percentage targets 0.5/1/1.5/2/3/5 against stops 0.5/1/1.5/2/3/5. Methods include full profit, activated trailing, breakeven, two partial-sale schedules (25% or 33% of original quantity twice, remainder trailed), momentum weakening (three falling closes, next-open exit), and current-rule replay. Detailed settings are in backtest-comparisons.csv. Current-rule replay may differ from recorded results because ticks, prior rule versions and execution assumptions differ.

The $10/$10 idea lost about $240.49 at 0.1% added cost per side in the low-first replay. Dollar targets are tiny relative to $3,000–$5,000 entries; on $5,000, $10 is just 0.2%, consumed by a 0.1% cost on each side before other friction. The most promising percentage alternatives turn negative under 0.5% per-side cost. That sensitivity prevents a live recommendation.

These are exit-only, in-sample counterfactuals using original entry times/sizes and original exit as fallback. They do not generate new entries, reallocate released cash or model order-book depth, minimum fees, partial fills, latency or market impact. Stops fill at the next observed open if gapped through; otherwise threshold fills within candles are optimistic for illiquid markets. Low-first/high-first are sensitivity assumptions, not rigorous bounds. Sparse candle gaps further limit evidence. Six deterministic edge-case tests check gaps, same-bar ambiguity, partial exits, breakeven costs, missing paths and boundary exclusion.

Top 3 strategies — research candidates only

Balanced: full exit +5%, stop -1%, no trailing, original recorded position sizes. Selected by net P/L minus half the closed-trade drawdown among the 84 plans under low-first and 0.1% cost; this is an explicit research preference, not a validated optimum. Conservative: identical exits with half every original position; figures scale linearly under the proportional-cost model only and are not an independent signal test. Higher-risk: full +5%, -1.5% stop, no trailing, original sizes; higher sample net accompanies larger drawdown. All three fail the heavier-cost stress. None establishes repeatable profitable weeks.

Recommended plan

Keep the proposed replacement in paper research. Exact first comparison: full +5% take profit, -1% stop, no trailing, compare original sizes with 50% sizes, costs at 0.1% and 0.5% each side, and retain the current rule as control. Collect timestamped bid/ask, fees, unique position identity, every candidate/entry/exit and unrealized peaks; evaluate later untouched weeks before any activation. This is a recommended experiment, not an instruction to change live settings. Tighter 0.5% stops did not solve the sample: TP5/SL0.5 lost $202.13 with $442.13 drawdown at 0.1% cost.

Weekly performance

Weeks use America/New_York ISO weeks of closure. Only weeks with observed closures enter the percentage; idle/unobserved weeks are not invented. Current week is partial. Robinhood spans years but has cost-basis gaps; demo and paper contain only one partial week. No proposed exit plan has enough weeks to establish weekly consistency. weekly-performance.csv includes week-end closed-equity drawdown; maximum within-week account drawdown remains unavailable without an equity time series.

Account-wide risk-control experiments

28 entry-filter scenarios test daily loss $100/$250, weekly $250/$500, position caps $500/$1,000, one slot, one per symbol, three-loss 30/60-minute pause, and 30/60-minute re-entry delay. These replay eligible closed-trade signals only; DEX ambiguous entries and still-open trades are excluded. Filters can skip signals but cannot invent their replacements. P/L scales with size, ignoring liquidity changes. Daily/weekly thresholds block later entries after realized losses; simultaneous positions and gaps can exceed them. No limit was activated.

A complete cross-broker exposure cap, correlated-asset cap and falling-asset re-entry rule cannot be credibly backtested without synchronized signals, equity and price histories. The concentration table establishes the need to research those controls, but no percentage was silently chosen or installed. Proposed threshold backtests on this one-week sample cannot certify portfolio safety.

Project inspection, scheduling and dashboard changes

The active project is on the Ubuntu server, with a local read-only audit mirror on this laptop. Inspected current source/log/state/configuration inventory across /root/trading_dashboard, /root/robinhood_bot, /root/etoro_bot, /root/etoro_live, /root/alpaca_bot, /root/crypto_bot and /root/dex_paper_bot. The collector retained 91 eligible files, excluded virtual environments, credentials and redundant backup copies, and recorded environment variable names rather than values. JSON/CSV/log files are the active stores discovered; no active trade database was identified in that inventory.

Scheduling includes Robinhood minute monitors, eToro real/Coinbase 15-minute checks, Alpaca weekday decision/monitor jobs, paper scanners/position monitors, Telegram notifier jobs and daily baseline/summary tasks. Repeated polling messages are not repeated fills. Logs include old DEX schema/key errors and API throttling; sampled monitoring and outages can miss a desired exit. Telegram notifier code was inspected but no message was sent by the audit.

Safe display changes: retain the earlier Coinbase credential-initialization fix and stale-baseline handling; add an audit report link and authenticated /audit view to the existing dashboard. This view shows all discovered accounts, explicit unknown values, exposure and performance. It is a dated snapshot, not a new live performance feed. It does not claim old DOGE tracker status represents all Robinhood holdings. Trading entrypoints, allocations, stops, targets, order cancellation and auto-trading configuration remain unchanged.

Files, validation and remaining work

New offline code: analyze.py, backtest.py, risk_replay.py, report.py and test_analysis.py under work/audit; reproducible research code is included in research-code.zip. New report/CSV outputs contain no account numbers or credentials. Dashboard deployment changes are recorded in validation.txt with the backup name, syntax checks, authentication tests and HTTP smoke results. No non-GET broker request was permitted by the account audit collector.

Remaining limitations: complete lifetime cost-basis and transfers; Robinhood equities outside this integration; broader Coinbase transaction history; continuous live/paper high-water marks; DEX token identity; a reconciled equity curve; multiple out-of-sample weeks; executable costs and liquidity; ongoing automated refresh of this audit snapshot. The older -$188.38 display cannot be interpreted as today’s trading result without its baseline/time. Do not promote the exploratory winner to live trading from this report.