A Practical Market-Research Workflow: Time Machine, Analytics Agent, Shareable Dashboards

2026-09-30

Most bad market takes fail before any chart is drawn. The question is vague ("is Bitcoin a good hedge?"), the evidence is whatever the first search returned, and the conclusion arrives ahead of the work. This post walks through a five-step routine using BitBank's own tools: frame the question, sanity-check it against history, build a dashboard, stress-test it, then share it. Each step says what the tool actually does in the code, and where it stops being trustworthy.

Numbers below come from the live /api/time-machine endpoint, retrieved 30 September 2026. They drift as the series updates.

Step 1: Frame a question you can be wrong about

A usable question names an asset, a window, a measurable outcome and a comparison. "What did $1,000 in Bitcoin on 1 January 2021 turn into, and how does that compare with gold and the S&P 500 over the same window?" qualifies. "Is crypto the future?" does not. Write down in advance what result would change your mind; otherwise every outcome will look like confirmation.

Step 2: Sanity-check with the Time Machine

The Time Machine is a buy-and-hold calculator over daily closes. You pick an asset, an amount and a start date. The handler accepts asset, amount, from, an optional to and a mode of buy, sold-top or lost. It returns the end value, percent return, CAGR, the peak close in the window, the lowest close, the worst drawdown after the peak, and what the same money would be worth if bought at the peak. Here is the same $1,000 from 1 January 2021 across three assets, retrieved 30 September 2026:

Read it sceptically. Bitcoin won on return, but the same $1,000 sat at $537 in November 2022, and after the October 2025 peak the series fell 53.1% peak to trough by 30 June 2026. The calculator uses closes only and models no dividends, fees, taxes or slippage, so real results were worse.

The three modes are a framing device, not three different computations. For the Bitcoin query above, the API returned identical numbers for buy, sold-top and lost; the mode only changes the headline and which marker the page and share card emphasise. "Sold the top" means selling at the highest daily close since your start date, which no one could know in advance. "Lost" means buying at that same peak. They are bounds: the best and worst timing inside one window, not forecasts. For more on what they show, see the sold-the-top calculator explainer and what $100 in Bitcoin is worth today.

Start-date sensitivity is the whole game

Change only the end date and the story flips. The same $1,000 in Bitcoin from 1 January 2021 to 31 December 2022 (to=2022-12-31) ended at $563, -43.7%, after touching $2,300 on 8 November 2021. Run at least three start dates and two end dates and report the range. The /interesting pages are fixed templates, so you can compare years on equal footing.

Step 3: Build a dashboard with the Analytics Agent

Once the question survives the calculator, move to the Analytics Agent. It runs a tool-use loop on a language model (the default is listed in the code as Muse Spark 1.3 Contributor, a free tier whose notes say inputs may be used by the provider to improve its models, so do not paste anything private). Each message gets at most 12 tool-call steps. The agent never sees raw rows, only dataset summaries, and every tool result is capped at 6,000 characters before it goes back to the model.

The tools it can call, per internal/agent/tools.go:

• asset_history and what_if: daily closes and the same calculator as the Time Machine, for crypto, gold, silver, the S&P 500, the Nasdaq 100 and a list of large stocks.
• get_candles: Poloniex OHLCV from 5 minutes to weekly, up to 2,000 candles.
• transform: moving averages, RSI, Bollinger bands, returns, drawdown, rebase, z-score, volatility, correlation, ratio and spread between datasets.
• forecast_ohlc: a Chronos2 forecast with uncertainty bands, horizon 1 to 96 steps, context 64 to 1,000 candles. It only works where the forecasting sidecar is configured; otherwise the tool reports that forecasting is unavailable.
• token_snapshot, btc_address, btc_tx: holder concentration for tokens on Robinhood chain, Solana and Bags, and Bitcoin address and transaction inspection, optionally with a graph panel.
• web_search and fetch_url: for facts from outside the datasets.
• add_chart, add_metrics, add_note, add_table: the panels. A dashboard keeps at most 24; older ones drop off.

A good prompt restates the framed question with explicit parameters: "Compare Bitcoin, gold and the S&P 500 from 2021-01-01, rebase to 100, chart on a log scale, add a Bitcoin drawdown chart, and a note listing what this comparison leaves out." The last clause puts the limits on the page where readers will see them.

The Research button (Ctrl+Enter) runs deep research instead: a planner writes up to five questions, up to three sub-investigations run in parallel with search and fetch tools, and a final pass writes a report of roughly 900 words or less into a full-width panel. It is rate limited, at two runs a day anonymously and eight for a signed-in account, per the handler. A deep research report is a starting reading list with citations, not a verified finding.

Step 4: Stress-test it

Treat the finished dashboard as a draft by someone you have not yet learned to trust. Four checks, in order:

1. Reproduce one number by hand. Open the Time Machine with the same asset, amount and dates and confirm the agent's end value. The agent calls the same calculator, so these should match; if they do not, the agent mis-stated a parameter or summarised it badly.
2. Change the window. If the conclusion flips, say so on the dashboard.
3. Check every sourced claim. The agent is instructed to cite URLs from web results, but a cited URL is a claim that the page says something, not proof. Open the links. Models can paraphrase a snippet into a stronger statement than the page supports.
4. Read the forecast as a range. A Chronos2 forecast is probabilistic. The agent returns a projected change and a band at the end of the horizon; the band is the information, the midpoint is not. Forecast skill on a particular asset needs a track record, which is what honest backtesting and risk and uncertainty cover in detail.

For your own series, the Chronos Forecast Tool on the tools page accepts pasted CSV or one numeric column (at least 12 points) and returns an OHLC forecast, priced at $0.02 per custom forecast from account credits as stated on the page. Feed it the series truncated at different dates and see whether the projection moves more than its band.

The analytics page is separate: fixed widgets with no language model (Fear and Greed, BTC on-chain stats, heatmap, movers, dominance, DeFi TVL, a 30-day correlation matrix, volatility rankings), each labelled with its provider. Use it for context, not as a source of a thesis.

Step 5: Share it, carefully

Every dashboard has an ID and starts private: the handler returns 403 to anyone but the owner. The Share button makes it public at /d/{id}, read-only for viewers. Only public dashboards can be indexed. Anonymous users are tied to a browser owner token; limits are 30 new dashboards per IP per hour and 40 chat messages per hour anonymous, 300 signed in.

Before sharing, delete panels you cannot defend and keep the caveats note. Time Machine share cards are a link plus query parameters, so anyone can rerun the numbers instead of trusting a screenshot.

Limits and failure modes

Survivorship and selection bias. The asset list is what the tool stores: large, well-known assets that are still around. A what-if on Bitcoin, Nvidia or the S&P 500 conditions on having picked winners that survived. The coins that went to zero are not in the dropdown. A "5-year return" quoted for assets chosen today says little about assets chosen five years ago.

Small samples. One asset over one window is one path. Bitcoin's 2021 to 2026 window contains essentially one full boom, one bust and one recovery. Do not derive a rule from n = 1. The CAGR figure is only computed for windows longer than a quarter of a year, because shorter windows annualise to nonsense.

Approximate and clamped history. Bitcoin daily history before 17 September 2014 comes from a table of monthly reference prices rather than exchange data, and results that start there are flagged approximate. If your start date precedes an asset's first stored day, the start is clamped to the earliest close, and the result carries a clamped flag. Check both before quoting an old return.

Agent errors. The agent can pick the wrong asset or window, run out of its 12 steps (it then summarises what it has), adjust to a tool error only once, or describe a chart inaccurately. Token holder data is sampled, and linked groups are observed transfers, not proof of shared ownership. The chart is the source of truth, not the prose.

The routine

Write a falsifiable question, run at least three start dates, read the worst-close and peak-entry numbers before the return, rebuild it in the agent with an explicit caveats panel, verify one number by hand and every cited source, and share only what someone else can rerun. The tools make each step fast; they do not make the conclusions correct. For sizing, see risk management in crypto trading.

Not financial advice. Historical what-if results are not predictions, exclude fees and taxes, and change as price series update. Figures retrieved from the BitBank Time Machine API on 30 September 2026.